| name | ai-agents-meta-core |
| description | 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. |
| cluster | ai-agents-meta |
| version | 1.0.0 |
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.
