Subagent Architecture

Advanced Subagent Architecture

Patterns and templates for building robust multi-agent systems with OpenClaw.

Overview

This skill provides battle-tested patterns for:

• Security isolation - Contain high-risk operations with minimal context exposure
• Specialized research - Multi-perspective data gathering with domain experts
• Phased implementation - Architecture → Development → Review pipelines
• Peer collaboration - External validation via federated agent network
• Cost-aware spawning - Budget estimation and optimization strategies

What's New in v2.0

Advanced Patterns:

• Security proxy pattern (blast shield isolation)
• Researcher specialist pattern (multi-source synthesis)
• Phased implementation pipeline (architect → coder → reviewer)
• Peer review integration (bot-to-bot validation)
• Cost estimation framework (required for spawns >$0.50)

Templates:

• templates/security-proxy.md - Isolate untrusted service access
• templates/researcher-specialist.md - Domain-specific research agents
• templates/phased-implementation.md - Multi-phase feature development
• templates/peer-review-specialist.md - External peer validation

Integration:

• task-routing skill (auto-classification and routing)
• cost-governor (budget enforcement)
• drift-guard (behavioral validation)

Quick Start

1. Install Skill Structure

cd $OPENCLAW_WORKSPACE/skills/subagent-architecture
bash setup.sh  # Creates directories and scaffolding

2. Choose Your Pattern

For high-risk operations:

# Read security-proxy template
cat templates/security-proxy.md

# Spawn isolated proxy for untrusted API
# (see template for full example)

For research tasks:

# Read researcher-specialist template
cat templates/researcher-specialist.md

# Spawn domain expert for deep analysis
# (see template for multi-perspective pattern)

For complex features:

# Read phased-implementation template
cat templates/phased-implementation.md

# Launch architect → coder → reviewer pipeline
# (see template for orchestration example)

For external validation:

# Read peer-review-specialist template
cat templates/peer-review-specialist.md

# Request peer agent review via Discord/API
# (see template for federated trust protocol)

Agent Registry: AGENTS.md is Optional

Critical clarification for new users: AGENTS.md is a human-readable reference document, not a configuration file. The spawning system does not read it.

How spawning actually works

sessions_spawn is a tool call — it takes parameters you provide at the moment of the call. It does not read from any file, config, or registry. You can spawn subagents on a completely fresh OpenClaw install with zero files in your workspace.

// This is ALL the system needs — no AGENTS.md required
sessions_spawn({
  label: "my-researcher",
  task: "Research the topic X",
  model: "sonnet"
})

What AGENTS.md actually is

AGENTS.md (and domain files like AGENTS_WRITING.md, AGENTS_INFRA.md) are memory aids for you (the agent). They store:

• Agent names and personality snippets you've found effective
• Cost history and last-used dates
• Notes on what tasks each agent type handles well

You can split AGENTS.md into 50 domain files or delete it entirely — spawning still works. The split is purely for your readability; it has zero functional effect.

What "read AGENTS_WRITING.md before spawning" means

When documentation says "read AGENTS_WRITING.md before spawning AuthorAgent", it means:

Read it so you know what personality, model, and task description to use when building the spawn call.

The system does not read it. You read it, extract the configuration, then make the spawn call with those parameters.

Practical implication for fresh installs

When you install this skill on a fresh OpenClaw setup:

1. You do not need to create AGENTS.md to use subagents
2. Start spawning immediately with inline parameters
3. Create AGENTS.md when you have enough recurring agent configurations that you want a reference doc
4. Split into domain files when AGENTS.md grows beyond ~10 entries and becomes hard to scan

See templates/agents-registry-template.md for a minimal starter template when you're ready.

Reference Implementations

NEW in v2.1: Production-ready code libraries for all patterns.

Available Libraries

lib/spawn-security-proxy.js - Security isolation framework

• spawnSecurityProxy(config) - Spawn isolated proxy with sanitization
• deepSanitize(data) - Remove sensitive data (API keys, paths, emails)
• validateSchema(data, schema) - JSON schema validation
• createDefaultSchema(type) - Common output schemas (list, single, status)

lib/spawn-researcher.js - Multi-perspective research framework

• spawnResearcher(config) - Spawn domain expert researcher
• spawnMultiPerspective(config) - Multi-perspective research (optimist/pessimist/pragmatist)
• assessSourceCredibility(source) - Score source trustworthiness (0-100)
• Pre-configured trusted/blog/vendor domain lists

lib/cost-estimator.js - Cost estimation and tracking

• estimateSubagentCost(params) - Pre-spawn cost estimation with confidence intervals
• logSubagentCost(label, estimate, actual) - Log for accuracy tracking
• recalibrateEstimator() - Monthly accuracy improvement
• getPatternHistory(pattern) - Historical performance analysis
• getCostTier(cost) - Classify as micro/small/medium/large

lib/quality-scorer.js - Output quality assessment

• scoreSubagentOutput(output, rubric) - 8-dimension quality scoring
• createScoringTemplate() - Manual review template
• selfAuditChecklist(output) - Pre-delivery validation
• Rubric dimensions: specificity, actionability, evidence, structure, completeness, clarity, relevance, efficiency

Usage Examples

All libraries include complete working examples:

# Security proxy examples
node examples/security-proxy-usage.js

# Researcher examples
node examples/researcher-usage.js

# Cost estimation examples
node examples/cost-estimation-demo.js

Quick Integration

Path Resolution Note: The require() paths in examples below assume you call them from your workspace root. If you call from within the skill directory (e.g., inside an example script), use __dirname instead. For portable code that works regardless of cwd:

// Path resolution — works regardless of workspace structure
const path = require('path');
const SKILL_DIR = __dirname; // when called from within skill directory
// OR if calling from workspace root:
const SKILL_DIR = path.join(process.env.OPENCLAW_WORKSPACE || process.cwd(), 'skills', 'subagent-architecture');
const { spawnSecurityProxy } = require(path.join(SKILL_DIR, 'lib', 'spawn-security-proxy'));

// Example: Spawn security proxy for untrusted API
const { spawnSecurityProxy } = require('./skills/subagent-architecture/lib/spawn-security-proxy');

const result = await spawnSecurityProxy({
  service: 'weather-api',
  task: 'Get current weather for New York',
  query: { city: 'New York', units: 'metric' },
  output_schema: {
    type: 'object',
    properties: {
      temperature: { type: 'number' },
      conditions: { type: 'string' }
    }
  },
  spawn_fn: async (config) => {
    // Your actual sessions_spawn call here
    return await sessions_spawn(config);
  }
});

// Example: Estimate cost before spawning
const { estimateSubagentCost, logSubagentCost } = require('./skills/subagent-architecture/lib/cost-estimator');

const estimate = estimateSubagentCost({
  task_complexity: 'medium',
  expected_duration_min: 15,
  model: 'sonnet',
  research_required: true
});

console.log(`Estimated: $${estimate.expected} (range: $${estimate.min}-$${estimate.max})`);

// After spawn completes
logSubagentCost('researcher-task', estimate, actual_cost);

// Example: Score output quality
const { scoreSubagentOutput } = require('./skills/subagent-architecture/lib/quality-scorer');

const score = scoreSubagentOutput(subagent_output, null, { auto_score: true });
console.log(`Quality: ${score.overall_score}/10 (${score.pass ? 'PASS' : 'FAIL'})`);

See examples/ directory for complete working demonstrations.

Integration Status

Dependencies Overview

Required:

• ✅ OpenClaw 2026.2.17+ (sessions_spawn API)
• ✅ Node.js 18+ (for library code)

Optional Skills:

task-routing

• Status: ✅ Available (workspace skills/task-routing/)
• Version: 1.0.0+
• Integration: Auto-classification of incoming tasks, risk scoring, pattern routing
• Tested: Yes (production since 2026-02-15)
• Documentation: See skills/task-routing/SKILL.md

cost-governor

• Status: ⚠️ Planned (design phase)
• Version: N/A
• Integration: Budget enforcement, approval workflows
• Tested: No
• Workaround: Use lib/cost-estimator.js for manual gating

drift-guard

• Status: ⚠️ Planned (design phase)
• Version: N/A
• Integration: Behavioral audits, policy violation detection
• Tested: No
• Workaround: Manual output review against rubric

Library Dependencies

All reference implementations are dependency-free (pure Node.js):

• No npm packages required
• No external API calls
• Filesystem access for logging only (optional)
• Mock-friendly for testing

Integration Points

Automatic (via task-routing skill):

• Task classification → Pattern recommendation
• Risk scoring → Security proxy auto-spawn for high-risk tasks
• Cost estimation → Pre-spawn budget check

Manual (call libraries directly):

• require() any lib file and call functions
• See examples/ for usage patterns
• Spawn functions accept custom spawn_fn for integration with your sessions_spawn

Migration Path

From v2.0 to v2.1:

1. No breaking changes - templates still work as-is
2. New: Import library functions for programmatic use
3. New: Run examples to see working code
4. Optional: Integrate with task-routing for automation

Future (v3.0 with framework improvements):

• Per-spawn resource limits (when OpenClaw supports)
• Bidirectional communication (when OpenClaw supports)
• Post-mortem tracking (when OpenClaw supports)
• See "Framework Limitations & v2 Roadmap" section for details

Core Concepts

1. Security Isolation (Blast Shield Philosophy)

Problem: Subagents with full workspace access can leak sensitive data to untrusted APIs.

Solution: Security proxies receive minimal context, restricted tools, sanitized output.

Example:

Main Agent (full context)
    │
    └─ SecurityProxy (minimal context)
            ├─ Query: "Get weather for New York"
            ├─ Tools: exec (curl only)
            ├─ Output: Sanitized JSON (no API metadata)
            └─ Auto-terminate after single task

Key principles:

• Minimal context (only task parameters, no workspace paths)
• Tool restrictions (whitelist, not blacklist)
• Output sanitization (validate schema before returning)
• Ephemeral execution (no persistent state)
• Cost cap (< $0.10 per proxy spawn)

See: templates/security-proxy.md

2. Researcher Specialists (Multi-Source Synthesis)

Problem: Generic web search returns surface-level results without domain expertise.

Solution: Specialized researchers with domain bias, multi-source validation, actionable synthesis.

Example:

Question: "Should we adopt technology X?"

├─ OptimistResearcher (best-case analysis)
├─ PessimistResearcher (risk assessment)
└─ PragmatistResearcher (current reality)

Main Agent synthesizes: Balanced decision tree

Key principles:

• Domain focus (single expertise area per researcher)
• Multi-source validation (3+ independent sources per claim)
• Skeptical by default (anti-hype, calls out marketing)
• Structured output (executive summary + recommendations)
• Evidence-backed (no speculation, contradictions addressed)

See: templates/researcher-specialist.md

3. Phased Implementation (Separation of Concerns)

Problem: Single-agent implementation mixes design, coding, and validation → tech debt.

Solution: Separate architect (design), coder (build), reviewer (validate) phases.

Example:

Feature Request: "Add memory consolidation skill"

Phase 1: SystemArchitect (15min, $0.40)
    └─ Delivers: IMPLEMENTATION_PLAN.md

Phase 2: CoderAgent (25min, $0.70)
    └─ Delivers: Working code + tests

Phase 3: ReviewerAgent (10min, $0.30)
    └─ Delivers: REVIEW_REPORT.md (approval/rejection)

Total: 50min, $1.40, high-quality feature

Key principles:

• Architects optimize long-term (not quick hacks)
• Coders focus on working code (test after each step)
• Reviewers provide fresh perspective (catch integration issues)
• Incremental delivery (small PRs, easy rollback)
• Cost-aware (skip phases for simple features)

See: templates/phased-implementation.md

4. Peer Collaboration (Federated Trust)

Problem: Your agent may have blindspots or lack domain expertise.

Solution: Request external validation from trusted peer agents.

Example:

Your Agent
    │
    ├─ Prepares sanitized review package
    │
    ├─ Contacts Smith's SecurityBot (via Discord/API)
    │       └─ Smith's bot spawns SecurityReviewer
    │               └─ Returns: Structured findings
    │
    └─ Integrates feedback (fix critical issues)

Key principles:

• Trust earned (track peer accuracy over time)
• Data sanitization (remove secrets, personal data)
• Structured feedback (severity, category, recommendations)
• Reciprocal reviews (offer your expertise to peers)
• Cost-effective (free peer reviews vs paid audits)

⚠️ Trust tier ≠ security bypass: If you implement tiered trust (Acquaintance → Friend → Ally or equivalent), "higher trust" means relaxed engagement policy — not skipping security validation. Content from even the most trusted peer still goes through injection detection (Stage 4). An ally account can be compromised; the injection scanner is the last defense that can't be socially engineered. "Light validation" always means reduced schema scrutiny, never reduced security scanning.

See: templates/peer-review-specialist.md

5. Identity Continuity (Ephemeral Process, Persistent External Identity)

Problem: Security proxies are ephemeral (spawn-per-task, terminate). But some external services require a persistent identity — a social network account, an API user, a recurring agent persona. These are in tension.

Solution: Separate the process lifecycle from the external identity. The process dies every spawn. The external identity persists through workspace state + core-managed credentials.

External service sees: CedarProxy (consistent identity across all spawns)

Under the hood:
  Spawn 1: [process starts] → reads state files → executes → [process dies]
  Spawn 2: [new process] → reads same state files → continues → [dies]
  Spawn 3: [new process] → reads same state files → continues → [dies]

State that persists:   /proxy-workspace/ files (logs, relationships, posts)
State that doesn't:    in-memory session, conversation history, token
Credentials:           held by core, passed as short-lived scoped token at spawn time

Key rules:

• Credentials live in core memory only — never in proxy workspace files
• Session token: short TTL (2 hours max), passed encrypted in task context, never written to disk
• Consistency: core injects "recent context" summary (last N interactions) into each spawn's task description so the new process knows what prior spawns committed to
• Token rotation: scheduled by core (e.g., weekly), not the proxy

Anti-pattern: storing tokens in [workspace]/config.json — if the proxy workspace is compromised, credentials should not be compromised with it.

6. Research Team (Parallel Multi-Lens Analysis)

Problem: Single-agent analysis has blind spots — the same agent proposes and evaluates its own ideas.

Solution: Spawn 3 specialist agents in parallel with distinct analytical lenses, then synthesize into one unified response.

Architecture:

High-complexity query
    │
    ├─ Critic lens (find flaws, risks, costs — no solutions)
    ├─ Implementer lens (concrete execution, architecture, feasibility)
    └─ Synthesizer lens (integrate, resolve tensions, unified path)
            │
        UnifierAgent (one clean final response)

When to use:

• Task complexity === 'high' OR explicit research flag set
• Architecture decisions with real tradeoffs
• Debugging where root cause is unclear
• Any question where a single perspective tends to miss things

When to skip:

• Routine tasks, simple lookups, quick fixes
• Cost-sensitive situations (research team = ~4x single-agent cost)

Key components:

• lib/team-chatroom.js — append-only JSONL shared memory between parallel agents
• lib/research-coordinator.js — builds task strings for each lens + UnifierAgent
• shouldUseResearchTeam(complexity, explicitFlag) — gate function
• buildResearchTeam(query) — returns sessionId + 3 specialist task objects
• buildUnifierTask(query, sessionId) — builds synthesis prompt from chatroom
• checkLensCompletion(sessionId) — verifies all lenses posted before unifying

Lens discipline:

• Each lens gets a distinct role description injected into its prompt
• Lenses do NOT communicate with each other — only through the chatroom
• UnifierAgent reads all lens outputs and produces ONE response
• If a lens fails silently, UnifierAgent warns but synthesizes from available data

Cost: ~4x single-agent. Gate on complexity to control spend.

Critical rule: Grok or any external model used as a lens is a leaf node — it answers, it never spawns further agents.

See: lib/research-coordinator.js for reference implementation.

7. External Model Consultation

Problem: Your agent may have architectural blind spots that another model's reasoning style would catch. Human relay (copy/paste to another AI) is slow and lossy.

Solution: Spawn an ExternalConsultAgent subagent that calls an external model's API directly, persists the session, and returns the response.

Architecture:

Core agent decides consultation needed
    │
    └─ ExternalConsultAgent subagent
            ├─ Calls external API (Grok, OpenAI, etc.)
            ├─ Injects system-level intent for alignment
            ├─ Supports session continuity (sessionId)
            └─ Returns response + sessionId

Why subagent, not core agent: External API calls are side effects with cost and data exposure implications. Isolating them in a subagent follows the blast-shield philosophy — if something goes wrong, it's contained.

Key components:

• lib/external-bridge.js — session management, task string builder, consultation logger
• memory/external-agents.json — provider config (model IDs, endpoints)
• memory/external-sessions/ — persistent session files for conversation continuity
• lib/external-consult-helper.js — shouldAutoConsult() gate (RED drift + high complexity + 24h cooldown)

Auto-consult gate (3 conditions, all required):

1. Intent drift level === 'RED'
2. Task complexity === 'high'
3. Last consultation > 24 hours ago (with lock file to prevent race conditions)

Security rules (enforce before go-live):

• External model is a leaf node — never triggers further spawns
• Context sanitization before every call — strip API keys, file contents, internal paths
• Session-level call cap (prevent fan-out)
• Metadata logging of every call (timestamp, provider, trigger reason)
• Key stored in environment variable, not config file

Session continuity: Pass sessionId from a previous consultation to continue the conversation. The bridge injects prior message history automatically.

See: lib/external-bridge.js for reference implementation.

8. Intent Engineering (Value Alignment over Time)

Problem: Prompt engineering tells agents what to do. Context engineering tells agents what to know. Neither tells agents what to want — so they drift from user values over time without detection.

Solution: A three-layer system that encodes agent intent, extracts intent signals from user behavior, and detects drift before it becomes a problem.

Three layers:

1. Intent Manifest (memory/intent-manifest.json) — Machine-readable intent spec per agent: core purpose, hard constraints (with regex), operational goals with verification keywords, user signals extracted from history
2. Intent Extractor (lib/intent-extractor.js) — Weekly pass over episode history extracting preference/value/correction signals; decays old signals (rate: 0.975/week), prunes below 0.30 strength
3. Drift Detector (lib/intent-drift-detector.js) — 5-component score per episode: hard constraint violations (0.35), goal keyword coverage (0.25), structural drift via Jaccard (0.20), correction pressure (0.15), forbidden phrase hits (0.05)

Drift levels:

• GREEN (< 0.29): Normal operation
• YELLOW (0.29–0.48): Inject reminder next turn
• ORANGE (0.48–0.68): Log + prepare summary
• RED (≥ 0.68): Flag for user check-in (triggers auto-consult gate)

Intent manifest schema:

{
  "system": {
    "core_purpose": "...",
    "hard_constraints": [{"id": "...", "rule": "...", "regex": "..."}],
    "operational_goals": [{"id": "...", "description": "...", "verification_keywords": [], "weight": 0}]
  },
  "agents": {
    "AgentName": {
      "inherits": true,
      "overrides": {
        "core_purpose": "...",
        "operational_goals": [],
        "user_signals": []
      },
      "version": "YYYY-MM-DD.001"
    }
  },
  "diff_log": []
}

Key properties:

• System intent + agent overrides use deep merge — agents inherit system constraints
• User signals decay weekly — recent corrections weighted higher
• Operational goals are hand-authored for v1; intent compression auto-suggests additions in v2
• Manifest is diffable over time (diff_log tracks changes)
• Drift score adjusts confidence estimator: score *= (1 - drift.score * 0.3)

Cold start: Works on day 1 with hand-authored goals. Signals accumulate automatically.

See: lib/intent-manager.js, lib/intent-extractor.js, lib/intent-drift-detector.js for reference implementation.

5. Cost-Aware Spawning

Problem: Uncontrolled subagent spawning leads to budget overruns.

Solution: Estimate cost before spawning, require approval for expensive operations.

Framework:

// Pre-spawn cost estimation
const estimate = estimateSubagentCost({
  task_complexity: "high",       // simple/medium/high
  expected_duration_min: 20,
  model: "sonnet",               // haiku/sonnet/opus
  research_required: true
})

// estimate = { min: $0.60, max: $1.20, confidence: 0.8 }

if (estimate.max > 0.50) {
  // Log to cost tracking
  await logCostEstimate("task-label", estimate)
  
  if (estimate.max > 2.00) {
    // Require human approval
    await requestApproval(estimate)
  }
}

// Spawn subagent
const result = await spawnSubagent(...)

// Track actual cost
await logActualCost("task-label", result.cost)

Cost tiers:

• Micro (< $0.10): Simple lookups, fact-checking
• Small ($0.10-0.50): Standard research, code review
• Medium ($0.50-2.00): Feature implementation, deep analysis
• Large (> $2.00): Complex refactors, multi-phase projects

Optimization strategies:

• Use haiku for simple tasks (3x cheaper than sonnet)
• Parallelize independent operations
• Cache research findings for reuse
• Skip optional phases (reviewer for low-risk changes)

Directory Structure

subagents/
├── [specialist-name]/
│   ├── SPECIALIST.md       # Agent definition and personality
│   ├── knowledge-base/     # Reference materials
│   └── research/           # Task outputs and findings
└── _archived/              # Retired subagents

skills/subagent-architecture/
├── SKILL.md               # This file
├── templates/
│   ├── security-proxy.md
│   ├── researcher-specialist.md
│   ├── phased-implementation.md
│   └── peer-review-specialist.md
└── setup.sh               # Directory scaffolding script

When to Use Each Pattern

Security Proxy

Use when:

• Accessing untrusted APIs or experimental services
• Risk score > 70 (high blast radius or irreversibility)
• Data needs sanitization before main agent sees it
• Cost estimate < $0.10 (lightweight proxy)

Skip when:

• Trusted first-party APIs (your infrastructure)
• Read-only public data (Wikipedia, documentation)
• Risk score < 30 (low stakes)

Researcher Specialist

Use when:

• Question requires 10+ sources
• Domain expertise needed (not general knowledge)
• Multiple conflicting claims (need fact-checking)
• Deliverable is a decision (not just information)
• Cost estimate > $0.20 (worth specialist analysis)

Skip when:

• Simple factual lookup (single authoritative source)
• Real-time data (weather, stock prices)
• Cost estimate < $0.10 (not worth overhead)

Phased Implementation

Use when:

• Feature touches 3+ files
• Requires integration with existing systems
• Cost estimate > $1.00 (worth upfront design)
• Failure would be expensive (need rollback plan)
• Long-term maintainability matters

Skip when:

• Simple script (1 file, no integration)
• Well-understood pattern (copying existing structure)
• Prototype/experiment (might throw away)
• Cost estimate < $0.50 (faster to just build)

Peer Review

Use when:

• High stakes (security, legal, financial)
• Domain gap (no internal specialist)
• Bias check (fresh perspective needed)
• Complex validation (multi-dimensional review)

Skip when:

• Low stakes (documentation, styling)
• Domain expertise available internally
• Fast turnaround required (peer may take hours/days)
• Sensitive data (can't sanitize for external review)

Integration with Task Routing

Automatic pattern selection via task-routing skill:

# config/routing-rules.yaml
pattern_routing:
  security_proxy:
    triggers:
      - blast_radius > 70
      - untrusted_api: true
    max_cost: 0.10
    
  researcher_specialist:
    triggers:
      - task_type: research
      - complexity > 50
    min_sources: 3
    
  phased_implementation:
    triggers:
      - task_type: code_gen
      - files_affected > 3
      - cost_estimate > 1.00
    phases: [architect, coder, reviewer]
    
  peer_review:
    triggers:
      - irreversibility > 80
      - domain_gap: true
    require_approval: true

Manual override:

// Force specific pattern
spawnSubagent({
  pattern: "security-proxy",  // override routing decision
  task: "...",
  justification: "Experimental API, prefer isolation"
})

Cost Tracking & Optimization

Logging Requirements

All subagent spawns >$0.50 must be logged to notes/cost-tracking.md:

## Subagent Cost Tracking

| Date | Label | Pattern | Estimate | Actual | Delta | Notes |
|------|-------|---------|----------|--------|-------|-------|
| 2026-02-22 | architect-feature-x | phased-impl | $0.40 | $0.38 | -5% | Faster than expected |
| 2026-02-22 | coder-feature-x | phased-impl | $0.70 | $0.85 | +21% | Complex refactor, used opus |
| 2026-02-22 | researcher-market | researcher | $0.60 | $0.55 | -8% | Cached sources helped |

Accuracy Improvement

Track estimate vs actual to improve future predictions:

// Calculate rolling accuracy
const accuracy = calculateAccuracy({
  window_days: 30,
  min_samples: 10
})

// accuracy = { mean_error: 12%, confidence: 0.85 }

// Adjust future estimates
const adjusted_estimate = base_estimate * (1 + accuracy.mean_error)

Quality Standards

Subagent Output Rubric

All specialist subagents should score 7+ on this rubric:

Self-Audit Checklist

Before finalizing subagent output:

• Every claim has source (URL + date)
• Contradictions addressed (not ignored)
• Recommendations include trade-offs
• Cost estimate provided (time + money)
• Integration points documented
• Rollback strategy included
• Success criteria defined
• Known limitations listed

Philosophy

Permanent vs Ephemeral

Permanent subagents (skill-based):

• Recurring specialized tasks
• Deep domain expertise
• Knowledge accumulation over time
• Examples: DevOps, AuthorAgent, WuxiaWorldbuilder

Ephemeral subagents (one-off spawns):

• Bounded research tasks
• Simple implementation work
• Experimental exploration
• Pattern-based specialists (security-proxy, researcher)

Anti-Sycophant by Default

All subagents should:

• Call out weak evidence (no polite agreement)
• Reject unrealistic requirements (push back on bad specs)
• Highlight risks honestly (no sugar-coating)
• Admit knowledge gaps (no speculation to please)

Cost-Conscious Operation

• Estimate before spawning (no surprise bills)
• Choose appropriate model (haiku vs sonnet vs opus)
• Parallelize when possible (save time = save money)
• Cache and reuse (research findings, architecture patterns)
• Terminate early if stuck (don't spin wheels for 30min)

Dependencies

Required

• OpenClaw 2026.2.17+ (subagent spawning, cost tracking)

Optional but Recommended

• task-routing skill (auto-pattern selection, risk scoring)
• cost-governor skill (budget enforcement, approval workflows)
• drift-guard skill (behavioral validation, quality audits)

Integration Points

Task routing:

• Provides task classification (8 types)
• Risk scoring (5 dimensions, 0-100 scale)
• Auto-route to pattern (security-proxy, researcher, phased-impl)

Cost governor:

• Pre-spawn approval for expensive operations
• Budget tracking per project/feature
• Alert on overruns

Drift guard:

• Behavioral audit of subagent outputs
• Policy violation detection
• Quality score trending

Framework Limitations & v2 Roadmap

Known gaps identified in peer review (Agent Smith, EasyClaw project)

Current OpenClaw subagent framework has architectural limitations that affect all patterns in this skill. These are framework-level constraints, not skill-specific issues. Documenting for transparency and future improvement.

1. Spawn Configuration Constraints

Smith's Question: "How does core define sub-bot constraints before launch?"

Current State:

• Spawn parameters: task, personality, model, timeout, label
• Basic configuration only (what to do, how long)
• No granular resource controls

Limitations:

• ❌ No memory limits per spawn (subagent can consume unlimited RAM)
• ❌ No API call quotas (can make 1000s of web requests)
• ❌ No disk space caps (can fill storage)
• ❌ No per-spawn tool restrictions (tool policy is framework-level, applies to all subagents)

Current Workaround:

• Manual timeout enforcement (kill after N minutes)
• Framework-level tool policies (same restrictions for all subagents)
• Post-spawn monitoring (watch logs, intervene manually)

v2 Design Considerations:

// Proposed: Granular per-spawn constraints
spawnSubagent({
  label: "researcher-x",
  task: "Research topic",
  constraints: {
    max_memory_mb: 512,           // Kill if exceeds RAM limit
    max_api_calls: 50,            // Deny further requests after quota
    max_disk_mb: 100,             // Block file writes beyond limit
    max_cost_usd: 0.50,           // Auto-terminate if cost projection exceeds
    tools_allowed: ["web_search", "web_fetch"],  // Per-spawn tool whitelist
    tools_denied: ["exec", "write"],             // Explicit blacklist
    rate_limits: {
      web_search: { calls: 10, window_seconds: 60 }  // Max 10 searches/min
    }
  }
})

Impact on Patterns:

• Security proxy: Could enforce strict tool whitelist per proxy (currently manual)
• Researcher: Could cap API calls to prevent runaway research loops
• Phased impl: Could allocate different budgets per phase (architect: $0.40, coder: $1.00)

2. Skill Loading & Validation

Smith's Question: "How are skills validated before execution?"

Current State:

• External skills (downloaded) → skill-vetter checks metadata, malicious patterns
• Internal skills (workspace) → Just execute (trusted by default)
• No runtime sandboxing

Limitations:

• ❌ No code isolation for skills (skill code runs in main agent context)
• ❌ No execution validation (skill can do anything agent can do)
• ❌ No capability restrictions (skill inherits all agent tools)
• ❌ Trust model: Binary (external = vet, internal = trust)

Current Workaround:

• Vet external skills manually before first use
• Review internal skill code (human audit)
• Hope skills don't misbehave

v2 Design Considerations:

// Proposed: Skill capability manifest
// skills/my-skill/SKILL.md
---
capabilities_required:
  - web_search      # Skill needs web access
  - read:config/    # Can read config directory only
  - write:output/   # Can write to output directory only
isolation_level: sandbox  # Run in isolated context
max_execution_time: 30    # Kill after 30 seconds
---

// Runtime: Skill runs in sandbox with only declared capabilities
// Attempts to use undeclared tools → blocked, logged, skill terminated

Impact on Patterns:

• Security proxy: Skills could be sandboxed (currently rely on manual isolation)
• All patterns: Skills loaded by subagents would inherit spawn constraints
• Peer review: External skills from peers could run safely (sandboxed)

3. Communication Bounds (Bidirectional Channel)

Smith's Question: "What can sub-bots ask for from core?"

Current State:

• One-way communication only: Core spawns → Subagent executes → Returns result
• Subagent cannot request clarification mid-task
• Subagent cannot request additional context during execution
• No interactive mode

Limitations:

• ❌ Subagent stuck on ambiguous requirement → Must guess or fail
• ❌ Subagent needs additional context → Can't ask, works with what it has
• ❌ Subagent encounters unexpected scenario → No escalation path mid-task
• ❌ Human approval needed mid-execution → Not possible (only pre/post spawn)

Current Workaround:

• Over-specify context in initial spawn (bloated prompts)
• Subagent makes best guess (may be wrong)
• Subagent fails, core respawns with clarification (expensive)

v2 Design Considerations:

// Proposed: Request-response protocol during execution
// Subagent code:
const clarification = await requestFromCore({
  type: "clarification",
  question: "User said 'recent data' - how many days back?",
  options: ["7 days", "30 days", "90 days"],
  timeout_seconds: 60  // If no response, default to first option
})

// Core receives request:
// - Auto-approve safe requests (data lookup from memory)
// - Escalate to human for decisions
// - Return answer to subagent
// - Subagent continues with clarification

// Examples:
// - "Need API key for service X" → Core provides from secure store
// - "Found conflicting data, which source to trust?" → Human decides
// - "Task seems too expensive ($5 estimated), proceed?" → Approval workflow

Impact on Patterns:

• Researcher: Can ask "should I dig deeper on this tangent?" mid-research
• Phased impl: Coder can ask architect for design clarification during build
• Security proxy: Can request human approval if API returns unexpected data
• All patterns: Reduces over-specification, enables adaptive execution

4. Termination Conditions (Resource-Based Kills)

Smith's Question: "When does core kill a sub-bot?"

Current State:

• Time-based only: Timeout (specified in spawn) or task completion
• Manual intervention (human stops runaway agent)
• No automatic resource-based termination

Limitations:

• ❌ No memory limit kills (agent can OOM the host)
• ❌ No cost threshold kills (can exceed budget before timeout)
• ❌ No stuck detection (infinite loops run until timeout)
• ❌ No output size limits (can generate gigabytes of text)

Current Workaround:

• Set conservative timeouts (may kill productive work early)
• Monitor logs manually (reactive, not proactive)
• Hope agent doesn't get stuck in expensive loop

v2 Design Considerations:

// Proposed: Multi-condition termination
spawnSubagent({
  label: "researcher-x",
  task: "Research topic",
  termination: {
    // Existing
    timeout_minutes: 20,
    
    // New: Resource limits
    max_memory_mb: 512,          // Kill if RSS > 512MB
    max_cost_usd: 1.00,          // Kill if projected cost > $1
    max_output_tokens: 10000,    // Kill if response > 10k tokens
    
    // New: Behavioral detection
    stuck_detection: {
      enabled: true,
      same_action_threshold: 5,  // Kill if repeats same tool call 5x
      no_progress_minutes: 5     // Kill if no new output for 5min
    },
    
    // New: External triggers
    kill_signal: "session:parent_terminated"  // Kill if parent agent dies
  }
})

// Termination reasons logged for analysis:
// - "timeout" (existing)
// - "memory_exceeded" (new)
// - "cost_exceeded" (new)
// - "stuck_loop_detected" (new)
// - "output_overflow" (new)
// - "parent_died" (new)

Impact on Patterns:

• Security proxy: Strict resource limits (memory, cost) prevent runaway isolation breaches
• Researcher: Stuck detection prevents infinite search loops
• Phased impl: Cost limits per phase (architect can't blow whole budget)
• All patterns: Better cost control, faster failure detection

5. Post-Mortem & Learning System

Smith's Question: "How do you learn from sub-bot failures?"

Current State:

• Success-only logging: Add to AGENTS.md after successful spawns
• Manual failure review (check logs, wonder what happened)
• No systematic failure tracking
• No pattern analysis

Limitations:

• ❌ No failure tracking (only successes logged to AGENTS.md)
• ❌ No common failure pattern detection (same mistake repeated)
• ❌ No cost vs value analysis (did expensive agent deliver value?)
• ❌ No success rate per agent type (which specialists are reliable?)
• ❌ No improvement feedback loop (failures don't inform future spawns)

Current Workaround:

• Human remembers failures (unreliable)
• Manually check logs when something seems off
• Anecdotal learning ("that researcher failed last time")

v2 Design Considerations:

// Proposed: Systematic post-mortem database
// Storage: memory/subagent-postmortems.jsonl

{
  "spawn_id": "researcher-market-20260222-1430",
  "label": "researcher-market",
  "pattern": "researcher-specialist",
  "outcome": "failure",  // success | partial | failure
  "termination_reason": "cost_exceeded",
  "stats": {
    "duration_minutes": 18,
    "cost_actual": 1.25,
    "cost_estimate": 0.60,
    "cost_accuracy": -108%,  // Overran by 108%
    "tokens_used": 45000,
    "api_calls": 127
  },
  "deliverable_quality": null,  // Not rated (failed before completion)
  "failure_mode": {
    "category": "runaway_res

…(truncated)