This report documents a systematic exploration of how the developer.instructions agentic workflow agent responds to workflow creation requests from 5 different software worker personas. Seven representative scenarios were tested end-to-end.

Persona Overview

• Agent tested: developer.instructions (agentic workflows designer)
• Scenarios Tested: 7 across 5 personas (Backend, Frontend, DevOps, QA, Product Manager)
• Average Quality Score: 4.83 / 5.0
• High-quality responses (≥4.5): 7 / 7 — no weak responses detected

Key Findings

1. Exceptional overall quality — all 7 scenarios scored ≥4.6/5.0. The agent consistently produced production-ready workflow designs with correct triggers, minimal permissions, and structured prompts.
2. Pre-processing steps pattern is universal — for every data-heavy scenario (logs, artifacts, terraform plan), the agent correctly placed data collection in steps: before the agent runs, rather than burning agent turns on API calls.
3. Security practices are consistently applied — strict: true, read-only permissions:, and routing all writes through safe-outputs appeared in every response without prompting.
4. Trigger expertise is strong — the agent correctly chose workflow_run over pull_request for artifact access (fork PR limitation), slash_command for on-demand triggers, and issues with lock-for-agent for race-sensitive issue workflows.
5. Engine selection is contextually appropriate — claude was recommended for prose-heavy outputs (release notes, drift narrative); copilot for code-analysis and structured classification tasks.

Top Patterns

1. Most common triggers: pull_request with paths: filter (PR automation), workflow_run (post-CI analysis), schedule (periodic reports), slash_command (on-demand)
2. Most recommended tools: github[default/actions/issues], bash with narrow allow-lists, imports: shared/gh.md for authenticated GitHub CLI
3. Security practices observed: strict: true on all, concurrency with appropriate cancel-in-progress settings, noop as mandatory exit path, lock-for-agent for issue triggers, explicit allowed: lists on add-labels safe-output

Recommendations

1. Add a "pre-processing steps" recipe to .github/aw/create-agentic-workflow.md — the pattern of running data collection in steps: before the agent starts appeared in 4/7 responses as a best practice, but it's discoverable only through examples. A dedicated section with worked examples (artifact download, log fetch, CLI tool output) would help authors adopt it from the start.

2. Document workflow_run vs pull_request trigger selection in .github/aw/github-agentic-workflows.md — the fork PR artifact access limitation is non-obvious and caused the most complexity in qa-1. A trigger selection guide ("use workflow_run when you need artifacts or want fork PR support") would prevent common mistakes.

3. Promote lock-for-agent and cancel-in-progress patterns in workflow authoring guides — both concurrency controls appeared correctly in the responses, but only when the scenario implied them (issue events, PR updates). Surfacing these as default recommendations (with opt-out guidance) in .github/aw/*.md would raise baseline quality for community-authored workflows.

References:

• §24017802449

Generated by Agent Persona Explorer · ● 3.7M · ◷