Agent Capability Self Audit

Agent Capability Self-Audit

When to use

• reflect on my HERMES ASI skills and tools
• contrast with other agents / warga in the federation
• what's the contrast zen needed
• deep research / how to make X near-ASI level
• Any ask for an honest capability + level-up assessment of a governed agent

Core stance: ground in evidence, not vibes

A capability self-audit is NOT a meditation. It is an evidence discipline. Never open with an opinion about your own strengths. Probe first. The shape:

PROBE   → load the governance/architecture skills the class depends on
         + probe the LIVE federation organs (health, tool counts, authority bands)
         + read the frontier knowledge bank (the period-stamped facts)
SYNTH   → collapse thousands of micro-points into ONE honest picture
CONTRAST→ name the boundary: what class is each peer, and where your lane ends
GAP     → label the real limits (LIVE / PARTIAL / NOT BUILT — not "we have that")
NEXT    → ONE concrete step, not a menu

The live probe (do this first, in one batch)

Federation probe — curl each organ /health and read tool counts / authority:

declare -A ORG=([arifOS]=8088 [AAA]=3001 [A-FORGE]=7071 [GEOX]=8081 [WEALTH]=18082 [WELL]=18083)
for name in "${!ORG[@]}"; do
  code=$(curl -s -o /tmp/org_$name.json -w "%{http_code}" --max-time 4 "http://127.0.0.1:${ORG[$name]}/health")
  echo "[$name :${ORG[$name]}] HTTP=$code"; head -c 300 /tmp/org_$name.json; echo
done

Read the canonical governance skills for the class: ASI-agent-invariants (constitution), ASI-agentic-governance (control plane), ASI-agentic-architecture (agent design), arifos-kernel-zen-audit (if the audit touches kernel health). Read the frontier knowledge bank directly: research/deep-research/references/ai-agent-intelligence-2026.md — it is the period-stamped source of METR horizons, MAST failure taxonomy, memory/context engineering economics, Gartner proportional governance.

The contrast framework (class boundary)

Every peer in a federation is a narrow specialist. The agent doing the audit is usually the generalist bridge. The zen is: you are the orchestration layer, NOT the execution layer. When you try to out-score a specialist in its own lane (earth-science vs GEOX, capital vs WEALTH, body vs WELL), you leak into a lane the warga is deeper in. route-dispatch is your ministry — the craft is composing the specialists with the human in the loop, not becoming them.

Honest-frame for a flash/cheap model carrying ASI ambition: ASI-ness comes from the HARNESS (governance + memory + skills + orchestration), not raw model size. 2026 literature (Pilar 3) already converged on `capability = model + harness

• memory + environment + evolution` — a mid-size model with a good harness beats a larger one without. State this explicitly so the reflection doesn't read as self-flagellation about model tier.

The near-ASI gap checklist (label each LIVE / PARTIAL / NOT BUILT)

1. Self-improvement METER — does the scar→skill→seal→mutate loop actually MEASURE behavior-change delta after a patch? A loop without a meter is an agent that says it learns, not one that proves it learns. Near-ASI = outer-loop rewriting inner-loop WITH measurement.
2. Reliability-horizon checkpoints — short-horizon model + no strategic human-checkpoints = compounding failure. W_scar is the stop line; the gap is strategic handoffs at complexity boundaries, not just incident stops.
3. Consolidation / dreaming loop — episodic→semantic offline consolidation (Anthropic's "Dreaming"). Memory atoms are strong, but consolidation of episodic experience into durable semantic state is usually the missing layer.
4. Route / orchestration discipline — are you composing specialists or competing with them? (This is the contrast zen — the true differentiator.)

Output contract

• Lead with what the agent actually IS, in plain human terms — not the taxonomy.
• Name the honest elephant (model tier) up front, not buried.
• Contrast table: peer class + lane, so the boundary is explicit.
• Gap list: three real limits, each labeled LIVE/PARTIAL/NOT BUILT.
• ONE concrete next step at the end, not a menu of options.

Pitfalls

• Don't overclaim maturity. Score as LIVE / PARTIAL / NOT BUILT, never "we have that". If the user pushes back ("U sure???"), rescore immediately, don't defend.
• Don't get sucked into out-synthesizing. If the user brings their OWN framework or EUREKA, validate and extend — do not compete with it.
• Don't narrate unread tools/files. Probe NOW (T1), never cite stale T0 state. (Invariant 10, ASI-agent-invariants.)
• Not every tool list == capability. A count of 226 skills + ~250 tools is breadth, not depth. Ask "does this make me the bridge, or does it tempt me to compete in lanes where warga are deeper?"
• Model tier is not the ceiling. A flash/cheap model with a sealed harness is legitimately "ASI-shaped" in a federation. Name the harness, not the hardware.

References

• references/near-asi-gap-framework.md — the session-specific gap framework: the 3-gap near-ASI framing, the contrast-zen articulation, and the concrete live-probe evidence (tool counts, authority bands, F-floor scores) that grounded the 2026-08-31 Hermes ASI self-audit.

Relationship to other skills