Free 2026 field manual

    Test whether your agent skills actually improve the work

    A practical manual for testing triggering, skill impact, tool use, outcomes, reliability, cost and security. Built from 221 sources published in 2026.

    Get the complete manual

    Build an eval suite you can trust

    Enter your email for the full tutorial, copyable eval schema, worked examples, CI cadence and source list.

    The direct answer

    A skill earns its place when it changes the result

    Test the same realistic task with the skill, without it and, when useful, with the wrong skill. Then inspect the route, deterministic contract, tool path, final state, repeated-run reliability, cost and security boundaries separately.

    Triggering and skill impact

    Measure whether the right skill activates and whether it improves the result against a controlled baseline.

    Contracts and trajectories

    Use code assertions for files, schemas, tool arguments, required steps, forbidden steps and final state.

    Reliability and cost

    Repeat stochastic cases. Report pass^k, confidence intervals, cost per success, latency and retries.

    Security and trust

    Test malicious instructions, unsafe actions, secret access, approval bypass, memory poisoning and retry duplication.

    Inside the manual

    From the first test case to a release gate

    • An eight-layer evaluation model
    • A decision table for choosing the right grader
    • A copyable JSON eval specification
    • Worked cases for routing, tools and side effects
    • Judge calibration and pass^k guidance
    • Security tests, CI cadence and a maturity model

    Research window

    Current 2026 evidence, with the limits stated

    The research combined papers, official engineering guidance, benchmark repositories and practitioner reports. Community sources informed failure discovery. They were not treated as product documentation.

    221
    sources reviewed
    205
    sources retained
    105
    core eval sources

    Evidence window: 1 January to 18 August 2026.

    Your next skill needs evidence, not confidence

    Start with 8 to 12 realistic cases, hard negatives, failure injection and three repeated runs. Expand only after the first suite finds real failures or protects a meaningful boundary.

    Preview the full manual

    The access page is indexed and available after signup too.