| name | agent-ergonomics |
| description | Evaluate and improve the Agent Experience (AX) of a tool, API, SDK, skill, or system — how well it serves an average agent. Use when the user says "agent ergonomics", "AX", "rate the agent ergonomics of X", "how agent-friendly is this", or invokes /agent-ergonomics. Both an evaluator (measure AX → a .work report) and a designer (prescribe AX fixes + patches). |
Agent Ergonomics (AX)
AX = Agent Experience. Measure and improve how well a system serves an
average agent — derived first-principles from what an agent fundamentally is
(perceives text + empirical queries/tests + sometimes images; acts only via tool
calls; bounded context; pays per token; little cross-call state; stochastic;
recursive, not looping; arrives with training priors; a human runs it).
This skill has two modes: evaluator (system → AX measurement) and designer
(measurement → prioritized fixes + ready-to-apply patches). Full rationale:
../../docs/ax-model.md.
The model — a matrix
Surfaces (rows) × orthogonal lenses (columns). A cell = how that surface
fares on that quality.
Surfaces: Disclosure (how the agent is taught) · Interface (calling it
correctly — incl. control mode: directness read-only→proposed→declarative→
relational→direct × ownership vendor↔fork × persistence sandboxed↔live) ·
Loop (operate turn-to-turn) · Recursion (the continual cross-turn
experience — memory, learning, progress, drift, compounding, state-machine ↔
memory alignment) · Human (the full collaboration/oversight relationship:
intent intake, observability, approval/escalation, steerability, trust/audit,
handoff, setup).
Lenses: Coherence (DRY↔WET) · Economy · Determinism ·
Verifiability (read and query/test) · Prior-alignment · Safety.
(Autonomy is a roll-up read across the matrix, not a column.)
Measurement — instrument toolbox (no fixed integer rubric)
Each cell is filled by whichever instrument(s) best capture it; validate the
instrument, not just the tool:
- Empirical trial telemetry · 2. Static/instrumented counters · 3. Perplexity /
cold-call accuracy · 4. Pairwise Elo / Bradley–Terry · 5. Item Response Theory
(latent trait) · 6. Trajectory / time-series · 7. Pareto frontier · 8.
Qualitative friction coding.
Combine into an AX profile (radar over surfaces×lenses) + Pareto placement
- optional latent IRT scalar. Never a number for its own sake.
Probe battery & test agent
Test agent: harness-agnostic runs use the Workbooks workagent standard
agent (washes out harness variance → measures the tool). Harness-specific runs
use that harness's architecture (and fetch its latest docs).
Comparability: task content is tool-specific; the measurement template is
constant. N = 5 trials/probe.
- T0 Cold-call — correct invocation with no docs (legibility, prior-alignment)
- T1 Guided task — representative artifact with docs (Loop, Disclosure, Economy)
- T2 Multi-turn project (Recursion, Coherence)
- T3 Failure-injection (Safety, recovery, failure-actionability)
- T4 Pairwise (holistic Elo)
- T5 Compaction survival — work → exhaust context → compact → continue, ≥3 levels; still advancing the goal? (Recursion)
- T6 Refinement / autonomy ceiling — how far past a first acceptable output; #completion states passed; self-review→query→correct→improve paths (autonomy, Verifiability)
- static — counters, no run.
Workflow
Screen → Deep → Designer.
- Screen (all subjects): static + T0 + one T1, single trial → first-pass
profile, flag agent-hostile cells.
- Deep (triaged subjects): full T0–T6, N=5, through ≥3 compaction levels →
full profile + Pareto (+ optional IRT).
- Designer: readings → prioritized fixes + ready-to-apply patches.
Output — a .work literate report
Always a .work report. Evidence = source blocks referencing real code at
real line numbers (cite live source, not paraphrase). Evidence standard:
any non-neutral cell reading (hostile/good/exemplary) must anchor to a source
block or a probe-run log; neutral/absent may be assertion-only. Reports are
written into the subject's own repo (e.g. out-of-thin-air/docs/ax/<tool>.work).
This project ships a standalone .work → markdown/HTML renderer.
Scale note
Cell readings are instrument outputs, not 1–5 scores. When a qualitative level
is unavoidable, name the state — Hostile (actively fights the agent) /
Absent / Tolerable / Good / Exemplary (amplifies the agent) — and always cite
evidence per the standard above.