Systematic exploration of how the agentic-workflows custom agent responds to workflow creation requests from 5 software worker personas. Tested 8 representative scenarios across PR automation, scheduled, and event-driven workflows.

Persona Overview

• Agent: agentic-workflows (developer.instructions)
• Scenarios Tested: 8 (from 5 personas × 2 tasks each)
• Average Quality Score: 4.6 / 5.0
• Workflow Run: §23993611929

Key Findings

• Perfect scores (5.0/5.0) on technically complex scenarios: DB migration review, vulnerability scanning, and deployment incident response — the agent excels when there is a well-defined taxonomy (risk categories, severity levels, root cause types)
• Pre-step pattern was independently recommended for 3 of 8 scenarios — downloading data before the AI session to save tokens is emerging as a strong convention
• repo-memory was correctly suggested for the only stateful scenario (vulnerability baseline diffing) — the agent understands cross-run persistence needs
• noop reminder was included in every generated prompt — 100% compliance on this critical requirement
• Security posture was consistently strong: network: defaults on all scenarios, comment/issue max: caps, and explicit fork-safety notes on PR automation

Top Patterns

1. Trigger selection was always appropriate — workflow_run + conclusion guard for incident response, push: tags with semver regex for releases, schedule + workflow_dispatch for all periodic work
2. Tool selection followed least-privilege — github: toolsets: [default] as the baseline, actions only when CI data needed, web-fetch only for external APIs (OSV, npm registry)
3. Safe-output type matched scenario purpose — create-issue for bugs/incidents, create-discussion for stakeholder digests, update-release for changelogs; close-older-issues: false on incidents, close-older-discussions: true on rolling weekly reports

Recommendations

1. Add repo-memory stateful scanning example to .github/aw/examples.md — the vulnerability scanner pattern (load baseline → scan → diff → write baseline) is broadly reusable for any "alert on new X" workflow and the agent independently converged on it
2. Document the pre-step data-fetching pattern in .github/aw/create-agentic-workflow.md — 3 of 8 scenarios benefited from downloading heavy data (logs, artifacts) before the AI session; this should be a first-class pattern with a template
3. Add a scheduled write-capable agent security checklist to .github/aw/github-agentic-workflows.md — scheduled agents with issues: write or contents: write are the highest-risk configuration; a concise checklist (network: defaults, max caps, noop required, concurrency group) would reduce misconfiguration

References:

• §23993611929

Generated by Agent Persona Explorer · ● 1.5M · ◷