Multi Agent Profile Redesign

Multi-Agent Profile Redesign

Design and implement a domain-based multi-profile Hermes fleet. Covers architecture, SOUL.md authoring at scale, Kanban coordination, Discord bot assignment, skill curation, and zero-downtime migration.

When To Use

WHENUSE: {NewFleetSetup,ProfileRedesign,MultiAgentArchitecture,FleetExpansion,MergeOverlappingProfiles}. ESPECIALLY:{ProfileCount>10,DomainOverlap,DiscordMultiBot}. NoSkip:{SingleProfile,SimpleRename}.

Core Design Principles

1. Domain-Based Profiles (Not Role-Based)

Organize profiles around skill domains, not job titles. Each profile owns a clear domain with minimal overlap.

❌ Role-based: coder, debugger, reviewer (overlapping)
✅ Domain-based: code (merged), creative, research, finance

Merge criteria: If two profiles share >40% of their skill catalog, merge them.

2. Kanban as Coordination Layer

Don't hand-design routing protocols. Use the built-in Kanban board:

• Board: Single main board with domain tags
• Dispatcher: Auto-spawns assigned workers on 60s tick
• Comments: Inter-profile communication on tasks
• Goal-mode: Loop workers until acceptance criteria met

Senna creates tasks → Kanban board → Dispatcher → Workers → Results flow back.

3. Orchestrator vs Worker Hierarchy

4. Bot Assignment Strategy

Not every profile needs a Discord bot. Decision matrix:

Typical: 6-10 bots for a 15-20 profile fleet.

SOUL.md at Scale

Template Hierarchy

Compressed DSL (Proteus-Style)

For 10+ profiles, use compressed DSL to keep token costs manageable. Average 68% reduction vs prose.

Sections to compress:

• IDENTITY → Trait.Trait.Trait. Name{Role}. CorePrinciple.
• PersRubric → NEO-PI-R 30 sub-facets in ~200 chars
• STYLE/AVOID/DEFAULTS → DSL encoding + arrow conditionals
• ROUTE → Structural compression: Design→creative|Build→code|Research→research
• TEAM → Token packing: {code:Impl,creative:Design,research:Data}

Sections to keep prose:

• Quality Gates (operational checklists)
• Section headers (human-parseable anchors)

Required Sections by Profile Type

DISCORD Section Pattern

Each bot-bearing profile needs:

DISCORD: Channel=#channel-name. {Role description}. When @mentioned, {behavior}. {Thread creation policy}. Completion: {output format}.

PersRubric Calibration by Role

Migration Strategy

9-Phase Zero-Downtime Migration

Phase 1: Create new profiles (--no-skills) — no disruption
Phase 2: Install SOUL.md files
Phase 3: Install skills (selective seed + pinning)
Phase 4: Configure models per profile
Phase 5: Discord restructure (channels + bots)
Phase 6: Initialize Kanban board
Phase 7: Migrate cron jobs
Phase 8: Testing (per-profile + integration + Discord)
Phase 9: Decommission old profiles (7-day hold)

Critical rule: Never remove old profiles until new ones are verified working. Keep backups at every phase.

Skill Curation for New Profiles

1. hermes profile create <name> --no-skills — start clean
2. Copy only relevant skills from source profiles
3. Evaluate external skill repos (Magnus, iknowkungfu, community) — see references/external-skill-repo-evaluation.md for the evaluation framework and case study
4. Pin critical skills: hermes --profile <name> curator pin <skill>
5. Let curator manage lifecycle (stale 30d → archive 90d)

Large Profile Handling (100+ skills)

For profiles with massive skill catalogs (cyber-red: 230, cyber-blue: 530):

• Use --no-skills + selective seeding
• Organize into subdirectories by sub-domain
• Agent sees only relevant subdirectory when navigating
• Pin top 5 most critical skills
• Consider splitting into sub-profiles if token budget exceeds limits

Skill Installation (Phase 3)

Phase 3 is the heaviest lift. Skills live in 3 locations (root, existing profiles, senna) and can be nested 2-3 levels deep. Use find to discover actual paths before writing copy commands — never trust the curation strategy document's paths verbatim.

For the full installation patterns, discovery scripts, batch copy approach, nested path gotchas, and Anthropic cybersecurity skills handling, see references/skill-installation-patterns.md.

Pitfalls

• --no-skills still installs bundled category directories. hermes profile create <name> --no-skills creates ~20-30 category subdirectories (apple, creative, software-development, etc.) with DESCRIPTION.md files but no actual skill content. This is expected — don't treat it as "something went wrong." The profile is ready for selective seeding.
• Skills can be nested 2-3 levels deep. The curation strategy may list mlops/llama-cpp but the actual path is mlops/inference/llama-cpp/. Common in mlops (inference/training/evaluation/models/research subdirectories). Always ls the category directory before writing copy commands. See references/skill-installation-patterns.md for the full gotcha table.
• Many skills exist ONLY in senna's profile, not in root. Senna accumulates skills from all domains. Skills like coding-size-limits, debug-artifact-cleanup, look-before-edit, pre-commit-security-checklist, and all 27 unreal-engine/ue-* skills are only in ~/.hermes/profiles/senna/skills/. When installing to new profiles, check senna as a source, not just root.
• Gateway can go down mid-redesign. Long multi-session work via Discord is vulnerable to gateway restarts (hermes update, crashes, manual shutdown). BEFORE any gateway-sensitive operation: (1) save all drafted files to filesystem, (2) save planning state to Mnemosyne + legacy memory, (3) write strategy/implementation docs to ~/.hermes/profiles/<coordinator>/cache/documents/. The Discord session is ephemeral — the files persist.
• Review Hermes docs before finalizing profile designs. New features (kanban, delegation roles, MCP filtering, --no-skills, curator pinning) may change the architecture. Pull docs from hermes-agent.nousresearch.com/docs and the hermes-agent skill before locking decisions. User explicitly said: "before we do anything, lets review the hermes agent docs completely to see whats all in hermes that could be newer than our last setup."
• Defer non-blocking decisions explicitly. UE5 server (separate vs section), voice channel (ElevenLabs TBD), cyber-blue split strategy — mark as [DEFERRED] with rationale, don't block the redesign on unresolved questions. User said: "if it flags issues then we will address it."
• Don't rename profiles while cron jobs reference them. Update cron first, then rename.
• Don't assume all profiles need a bot. Most workers are Senna-subordinated.
• Don't over-plan. Start, flag issues, address them. "If it flags issues then we will address it."
• Don't skip the backup step. Always create a full backup before migration phases.
• When delegating bulk SOUL.md drafting, lock format spec in the prompt. Subagents will inject inconsistencies: PersRubric = vs : delimiters, sub-profile routing in merged profiles, legacy name references. Always specify: (1) PersRubric uses : and spaces, (2) merged profiles don't route to sub-profiles that no longer exist, (3) team rosters use new profile names only. Run a consistency pass after batch delivery.
• read_file in execute_code deduplicates. If you read a file, then write it, then try to read it again in the same execute_code script, read_file returns content_returned: false with a "file unchanged" message. Use terminal("cat <path>") instead for re-reads within scripts.
• find output can be garbled in execute_code. The find command's output sometimes gets mangled when run inside execute_code scripts — line counts appear as 1F 1D: prefixes instead of clean paths. Use terminal() directly for discovery commands.
• maxdepth filter affects skill counts. When counting skills in cyber-red/cyber-blue, find -maxdepth 2 shows only 6-7 directories because the Anthropic skills are nested deeper. Use maxdepth 1 on the Anthropic skills subdirectory itself to get the real count (755).
• Cron jobs use channel IDs, not profile names. The deliver field in cron jobs uses Discord channel IDs (e.g., discord:<id>), not profile names. Changing profile names or assignments doesn't break cron delivery. This contradicts earlier risk assessments — the actual migration risk for cron is none.
• Discord admin requires gateway session. The discord_admin toolset is only available through the Discord gateway adapter where the bot token is accessible. CLI sessions can't access the bot token. Workaround: prepare commands for user to paste in Discord where the gateway has admin access.
• Don't forget cross-profile memory design. Mnemosyne private vs shared, Fabric, kanban comments — each has different visibility.
• Token budget matters. cyber-blue with 530 skills will blow up context. Use progressive disclosure + subdirectory organization.
• Model assignment per profile. Not all profiles need the same model. Use cheaper models for lightweight workers.
• The user wants Senna as single reporting channel. All specialist output flows through Senna. Don't set up direct user-to-worker paths unless explicitly asked.
• hermes profile create seeds builtin skills but NOT config.yaml or .env. The create command gives you SOUL.md and category directories with builtin skill stubs, but config.yaml and .env must be written manually. If hermes profile list shows a profile with a default model, that's Hermes' fallback — not a real config. Always write config.yaml explicitly. See profile-bootstrapping skill's references/batch-config-generation.md for the batch generation pattern.
• Hermes has an internal secrets store separate from .env files. API keys stored via hermes config set or the setup wizard are visible via hermes config show but not readable as plaintext. Profiles may access these keys without .env files. Don't block on missing .env — verify with a smoke test instead.
• Nous Portal free tier (deepseek-v4-flash:free) is no longer available (as of June 2026). Use OpenRouter free models instead. OpenRouter has 26 free models including Owl Alpha (1M context, tools) and DeepSeek V4 Flash. Rate limits: 20 req/min, 200 req/day per model.
• Magnus skills come from a separate git repo. git.brandyapple.com/magnus/agent-skills — not in the iknowkungfu registry. Clone, then copy to profile's skills/magnus/ directory. See profile-bootstrapping skill's references/magnus-skill-installation.md for the full inventory and installation pattern.

References

See references/external-skill-repo-evaluation.md for the evaluation framework when assessing third-party skill repos (Magnus, iknowkungfu, community), including the Magnus case study with per-profile installation results and Windows team applicability matrix.

See references/batch-soul-drafting-pitfalls.md for pitfalls and fixes when delegating bulk SOUL.md drafting to subagents — format inconsistency, legacy names, orchestrator vs worker sections, and the post-delivery consistency pass.

See references/skill-installation-patterns.md for the actual skill discovery process, batch installation scripts, nested path gotchas, and Anthropic cybersecurity skills handling.

See references/gateway-interruption-recovery.md for the save-before-restart checklist and recovery playbook when the gateway goes down mid-redesign — what to save, where, and how to resume.

See the hermes-soul-authoring skill for SOUL.md templates, PersRubric calibration, compressed DSL techniques, and the 16-point critical review checklist.