Agent-useful products

Start with real tasks, not a checklist of fashionable protocols. Retain the core method: action parity, context parity, stable object identity, bounded outputs, verification, and recovery. A documented capability is not an implemented one; a directory of manifests is not evidence that an agent can finish the task.

Audit an existing surface

Select 5–10 representative tasks with expected outcomes, including a read, an authorised mutation, a denied operation, and interrupted-work recovery.
Map each task's necessary context, object IDs, available action, permission, success evidence, and retry/resume path. Use the noun test: discover, identify, read, change where appropriate, verify, recover.
Inspect source/contracts, then execute only the authorised checks. Record whether each observation is a source signal, documented claim, or runtime proof.
Compare the human and agent paths. Preserve intentionally human-only consent, MFA, CAPTCHA, and biometric boundaries; do not score bypassing them as parity.
Prioritise demonstrated blockers and test the proposed repair on the same tasks. Report unresolved evidence rather than turning missing tool access into a failure.

Use the report template, architecture methods, and evaluation guidance as needed. The retained local/web scanners and scoring rubric are heuristic inventory aids, not protocol validators or measured agent-success scores. A missing optional discovery convention is not an interoperability defect without a consumer that actually requires it.

Design the smallest useful contract

Prefer the surface already native to the product. A small CLI may need help, structured output, and explicit exit codes, not an A2A server. An HTTP integration may need its current OpenAPI contract and SDK, not a parallel handwritten client. Use a workflow-level operation when it owns atomicity or a real server-side job; otherwise expose composable operations with identifiable results.

For writes, define what happens after timeout, partial success, concurrent edits, and repeated delivery. Request IDs are not automatically idempotency keys. Keep confirmed outcomes separate from unknown outcomes, and define the retention/scope of any deduplication guarantee. A preview or hash cannot authorise additional work. Long-running tasks need durable IDs, explicit terminal states, and bounded waits; local cancellation is not proof remote work stopped.

Read web/discovery contracts or API/MCP/A2A contracts for the chosen surface. The A2A example now targets protocol 1.0; /.well-known/mcp.json and the bundled skills index are local design examples, not universal standards. A2A skills are protocol capability descriptions, not filesystem Agent Skills packages.

Inspect a skill without executing it

Resolve SKILL_DIR to the installed directory containing this file. Paths below are skill-relative, not scripts expected in the target repository.

python "$SKILL_DIR/scripts/validate_agent_assets.py" --skill-dir /absolute/path/to/skill

Install the declared PyYAML dependency in an authorised environment first, or run the implementation file with a PEP 723-aware runner. Validation parses real YAML, checks required fields, JSON/Python syntax, and conventional relative Markdown links. It does not execute scripts, write bytecode, contact services, validate complete protocol schemas, or evaluate whether the skill improves agent behaviour. The old --run-help and --py-compile modes are removed. Running even --help can execute arbitrary project code; do that separately only after trust review.

See skill/instruction checks. Optional directories are optional, not automatic quality failures. Prefer current commands and useful examples rather than retaining aliases solely for compatibility.

Generate drafts, not deployed discovery endpoints

python "$SKILL_DIR/scripts/generate_agent_assets.py" \
  --output /private/existing-parent/new-drafts --project-name "Example" \
  --base-url https://example.com --surface a2a

This previews without creating directories. Select each surface explicitly; add --write only to create a new private draft directory. Existing paths are never overwritten. The final manifest marks completion; an interrupted write can leave a partial directory. No --force, default all, live .well-known publication, or automatic replacement of a project's AGENTS.md remains.

Review generated placeholders and validate the protocol with the actual chosen SDK/schema before publication. Declare only implemented capabilities and enforced authentication. The A2A fixture uses HTTP+JSON, a real service endpoint separate from the card URL, and a bearer-auth declaration that the server must implement.

Maintenance evidence

Run python -m unittest discover -s "$SKILL_DIR/tests" -v for local regression checks. The new fixture tests cover scaffold/validator behaviour and selected A2A shape regressions, not full A2A or MCP conformance. Preserve the substantive domain references, evaluation seeds, capability maps, and existing MIT licence.

Reviewed 2026-09-13 against Agent Skills, authoring guidance, A2A 1.0, and RFC 9727.

agent-use