AI Agents Meta Core
Shared model for the ai-agents-meta cluster. The planning, orchestration, loop, audit, and
cost spokes all depend on these interlocking ideas — keep them consistent here so no spoke
contradicts another.
1. The decision this cluster turns on: eval-first × default-deny autonomy
Every spoke is an answer to one question: how much can this agent do on its own, and how do
we know it did it right? The cluster's stance is two coupled rules:
Plan ──gated by──> Eval ──authorizes──> Autonomous step ──bounded by──> Tool / loop budget
- Eval-first — no step that an agent runs unattended ships without a gate that can fail
it (a test, a review, a schema/assertion, a screenshot diff). The gate is designed before
the step, not bolted on after. →
agentic-engineering, dynamic-workflow-mode, continuous-agent-loop
- Default-deny autonomy — an agent gets the narrowest tool set, permission, and loop
budget that lets the gated step pass. Widening any of them (a new tool, a higher iteration
cap, write access) is a governance change worth stating out loud. →
team-agent-orchestration, agentic-os
Rule: if you can't name the gate that fails the step, the step isn't ready to be
autonomous — drop it back to a human checkpoint.
2. The agent stack (what the audit spokes inspect)
Agent failures localize to a layer. Reason top-down before patching:
| Layer (top → bottom) |
Typical failure |
Spoke |
| Prompt / intent |
vague ask, missing constraints |
prompt-optimizer |
| Plan / decomposition |
steps too big, no cold-start brief |
blueprint, plan-orchestrate |
| Orchestration |
ownership unclear, merge collisions |
team-agent-orchestration |
| Loop / control |
runs forever, no recovery |
continuous-agent-loop |
| Tool discipline |
wrong tool, hidden repair loops |
agent-architecture-audit |
| Memory / state |
pollution, cross-project bleed |
agentic-os, continuous-learning-v2 |
| Model / cost |
over-powered model for trivial step |
cost-aware-llm-pipeline, token-budget-advisor |
When something breaks: agent-architecture-audit localizes the layer →
agent-introspection-debugging runs the contained recovery.
3. Shared conventions
- Research before building. Every build path opens with
search-first — find the existing
tool/lib/skill before hand-rolling one. The cheapest agent step is the one you didn't write.
- Cold-start briefs. A planned step must be executable by a fresh agent with no prior
context — self-contained inputs, outputs, and done-criteria. →
blueprint.
- Generative, not executive, advisors.
prompt-optimizer and plan-orchestrate emit
paste-ready artifacts; they never run the task themselves. Keep that boundary.
- Project-scoped learning. Instincts/skills minted from one project stay scoped to it;
never let learned state contaminate another repo. →
continuous-learning-v2.
4. Cost-routing tiers
Route by task complexity, not habit:
| Task shape |
Route |
Spoke |
| Deterministic / structured |
code or a small/cheap model |
cost-aware-llm-pipeline |
| Bounded reasoning, repeatable |
mid-tier model + cache |
cost-aware-llm-pipeline |
| Open-ended / high-stakes |
top model, gated |
agentic-engineering |
| Per-answer depth control |
offer the user a budget |
token-budget-advisor |
Prompt caching, retry-with-backoff, and budget tracking are baseline, not optional →
cost-aware-llm-pipeline.
5. Gate / loop / cost matrix
| Concern |
Default posture |
Widen only when… |
| Autonomy |
human checkpoint at each handoff |
the step has a failing-capable gate |
| Tools granted |
read-only, minimal set |
the gated step provably needs more |
| Loop budget |
hard iteration + time cap |
recovery + eval prove convergence |
| Model tier |
cheapest that passes the gate |
complexity demands it |
6. Version / tooling
- Built for Claude Code as the primary host (slash commands, skills, hooks, file-based
memory) but the patterns are host-agnostic. →
agentic-os, dynamic-workflow-mode.
continuous-learning-v2 is v2.1 (project-scoped instincts; supersedes v1's global store).- Orchestrators that namespace agent/command names by install form keep one form per output
— never mix plugin-prefixed and bare names. →
plan-orchestrate.
7. Shared guardrails
- Eval-first: no autonomous step without a gate that can fail it.
- Default-deny autonomy: narrowest tool/permission/loop budget that works; state every widening.
- Search before build:
search-first opens every build path.
- Route by complexity: cheapest model/tier that passes the gate; caching + retries baseline.
- Human at the handoff: every orchestration boundary keeps a decision point.
- Generative advisors stay advisory: prompt/plan emitters never self-execute.
- Sibling cluster: this cluster owns eval-first discipline; for the eval harness / agent
benchmarking tooling itself →
quality-eval cluster (agent-eval, eval-harness,
verification-loop, benchmark, production-audit).
- This cluster is the meta layer — for a concrete domain agent (scraping, trading, a product
feature), build it with these patterns rather than reaching for a one-off domain skill.
1---2name: ai-agents-meta-core3description: Shared reference for the ai-agents-meta cluster: the eval-first + default-deny-autonomy model every spoke turns on, plus shared conventions (the agent stack, cold-start briefs, cost routing tiers) and the gate/loop/cost matrix. USE WHEN wiring an autonomous loop, granting an agent tools, setting an eval gate, or routing by model cost — the interlocking rules every meta spoke shares.4---56# AI Agents Meta Core78Shared model for the `ai-agents-meta` cluster. The planning, orchestration, loop, audit, and9cost spokes all depend on these interlocking ideas — keep them consistent here so no spoke10contradicts another.1112## 1. The decision this cluster turns on: eval-first × default-deny autonomy1314Every spoke is an answer to one question: **how much can this agent do on its own, and how do15we know it did it right?** The cluster's stance is two coupled rules:1617```18Plan ──gated by──> Eval ──authorizes──> Autonomous step ──bounded by──> Tool / loop budget19```2021- **Eval-first** — no step that an agent runs unattended ships without a *gate that can fail22 it* (a test, a review, a schema/assertion, a screenshot diff). The gate is designed *before*23 the step, not bolted on after. → `agentic-engineering`, `dynamic-workflow-mode`, `continuous-agent-loop`24- **Default-deny autonomy** — an agent gets the **narrowest** tool set, permission, and loop25 budget that lets the gated step pass. Widening any of them (a new tool, a higher iteration26 cap, write access) is a governance change worth stating out loud. → `team-agent-orchestration`, `agentic-os`2728**Rule:** if you can't name the gate that fails the step, the step isn't ready to be29autonomous — drop it back to a human checkpoint.3031## 2. The agent stack (what the audit spokes inspect)3233Agent failures localize to a layer. Reason top-down before patching:3435| Layer (top → bottom) | Typical failure | Spoke |36|---|---|---|37| Prompt / intent | vague ask, missing constraints | `prompt-optimizer` |38| Plan / decomposition | steps too big, no cold-start brief | `blueprint`, `plan-orchestrate` |39| Orchestration | ownership unclear, merge collisions | `team-agent-orchestration` |40| Loop / control | runs forever, no recovery | `continuous-agent-loop` |41| Tool discipline | wrong tool, hidden repair loops | `agent-architecture-audit` |42| Memory / state | pollution, cross-project bleed | `agentic-os`, `continuous-learning-v2` |43| Model / cost | over-powered model for trivial step | `cost-aware-llm-pipeline`, `token-budget-advisor` |4445When something breaks: `agent-architecture-audit` localizes the layer →46`agent-introspection-debugging` runs the contained recovery.4748## 3. Shared conventions4950- **Research before building.** Every build path opens with `search-first` — find the existing51 tool/lib/skill before hand-rolling one. The cheapest agent step is the one you didn't write.52- **Cold-start briefs.** A planned step must be executable by a *fresh* agent with no prior53 context — self-contained inputs, outputs, and done-criteria. → `blueprint`.54- **Generative, not executive, advisors.** `prompt-optimizer` and `plan-orchestrate` *emit*55 paste-ready artifacts; they never run the task themselves. Keep that boundary.56- **Project-scoped learning.** Instincts/skills minted from one project stay scoped to it;57 never let learned state contaminate another repo. → `continuous-learning-v2`.5859## 4. Cost-routing tiers6061Route by **task complexity, not habit**:6263| Task shape | Route | Spoke |64|---|---|---|65| Deterministic / structured | code or a small/cheap model | `cost-aware-llm-pipeline` |66| Bounded reasoning, repeatable | mid-tier model + cache | `cost-aware-llm-pipeline` |67| Open-ended / high-stakes | top model, gated | `agentic-engineering` |68| Per-answer depth control | offer the user a budget | `token-budget-advisor` |6970Prompt caching, retry-with-backoff, and budget tracking are baseline, not optional →71`cost-aware-llm-pipeline`.7273## 5. Gate / loop / cost matrix7475| Concern | Default posture | Widen only when… |76|---|---|---|77| Autonomy | human checkpoint at each handoff | the step has a failing-capable gate |78| Tools granted | read-only, minimal set | the gated step provably needs more |79| Loop budget | hard iteration + time cap | recovery + eval prove convergence |80| Model tier | cheapest that passes the gate | complexity demands it |8182## 6. Version / tooling8384- Built for **Claude Code** as the primary host (slash commands, skills, hooks, file-based85 memory) but the patterns are host-agnostic. → `agentic-os`, `dynamic-workflow-mode`.86- `continuous-learning-v2` is **v2.1** (project-scoped instincts; supersedes v1's global store).87- Orchestrators that namespace agent/command names by install form keep **one form per output**88 — never mix plugin-prefixed and bare names. → `plan-orchestrate`.8990## 7. Shared guardrails9192- **Eval-first**: no autonomous step without a gate that can fail it.93- **Default-deny autonomy**: narrowest tool/permission/loop budget that works; state every widening.94- **Search before build**: `search-first` opens every build path.95- **Route by complexity**: cheapest model/tier that passes the gate; caching + retries baseline.96- **Human at the handoff**: every orchestration boundary keeps a decision point.97- **Generative advisors stay advisory**: prompt/plan emitters never self-execute.98- **Sibling cluster**: this cluster owns eval-first *discipline*; for the eval harness / agent99 benchmarking *tooling* itself → `quality-eval` cluster (`agent-eval`, `eval-harness`,100 `verification-loop`, `benchmark`, `production-audit`).101- This cluster is the **meta** layer — for a concrete domain agent (scraping, trading, a product102 feature), build it *with* these patterns rather than reaching for a one-off domain skill.
Sheshiyer