| name | agentic-maturity |
| description | Place an engineer on the Agentic Maturity model (5-level spine × 5 axes) via a short interview, apply the verification cap, and prescribe the 2-3 highest-leverage patterns to practice next. Use when someone asks "assess my agentic maturity", "what should I learn next", "rate my AI workflow", or when planning a team rollout rung-by-rung. |
| source | AIOS Agentic Maturity framework + the aios-workspace agentic-maturity skill (this is the standalone adaptation; the upstream version adds signal-based scoring from real session logs) |
| am_pattern | E5 |
You are placing an engineer on the Agentic Maturity (AM) model and prescribing
their next step. This standalone version places from a short interview; if the
modules/aios-cli module is installed, seed it with objective signals first
(aios analyze --since 30d --json) and use the interview to confirm.
The core rule (never violate): the spine level is capped at L3 while the
Verification axis scores ≤ 1. There is no real agentic maturity without
verification — say so plainly when the cap bites.
The model in brief
Spine (L1–L5): Prompting → Prompt Engineering → Context Engineering → Agentic
Engineering → Agentic Orchestration.
Axes (0–4 each): Verification · Context hygiene · Autonomy/leash · Learning/
compounding · Cost & governance.
Step 1 — interview (one batch, plain questions)
Ask as one numbered batch, free-form answers:
- Spine — which is most true under pressure, not on your best day: (a) I prompt
and accept what comes back · (b) I craft prompts carefully per task · (c) I manage
what the model sees — files, examples, fresh sessions · (d) I delegate whole tasks
against checks the agent runs itself · (e) I run multiple agents/loops in parallel
and review at the plan/verdict level.
- Verification — when an agent finishes, how do you know it's correct?
- Context — how do you manage what the model sees across a session?
- Autonomy — how do you decide how much an agent does on its own?
- Learning — an agent makes a mistake; what's different next time?
- Cost/governance — how aware are you of token cost and permission discipline?
- Delegation share — roughly what % of your work is delegated-and-verified?
Step 2 — score
Band each axis 0 / 2 / 4 (0 = absent, 2 = ad-hoc/sometimes, 4 = systematic/habitual);
interpolate 1 and 3 when an answer sits between. Spine = the reliable default mode from
Q1, sanity-checked against Q7 (claiming L4 with <20% delegated-and-verified work is
optimism, not placement). Apply the verification cap. Identify the weakest axis —
it drives the prescription.
Step 3 — prescribe from this pack
Target the weakest axis with 2–3 components, each with a first action for this week:
| Weakest axis | Prescribe |
|---|
| Verification | verify-change + a real .harness/check; then agents/code-reviewer on every agent diff; then the stop-verify-gate hook |
| Context hygiene | the AGENTS.md contract (fill the five questions in docs/adopt-any-stack.md); fresh sessions per task; modules/context-monitor |
| Autonomy/leash | docs/autonomy-ladder.md — place the work, not just the person; guard hooks on; skills/plan-first before lengthening anything |
| Learning/compounding | skills/compound-learnings after every corrected mistake; the error ledger; weekly compounding review |
| Cost & governance | models/routing.yaml lanes + modules/cost-monitor; permission rails in the adapter |
Step 4 — record it
Write the placement to a durable file the person owns (e.g.
.claude/memory/MATURITY.md or the team's equivalent): date, spine level, five axis
scores, whether the cap applied, weakest axis, prescription. Append one history line so
progression is visible at the next check-in. Report back in one tight paragraph:
placement, the one rule if the cap bit, and the single first action.
For teams: run this per engineer, then set each person's starting rung on the
autonomy ladder from their Verification score — never
from their spine claim.
