Agent Persona Exploration - 2026-05-08

[github-actions[bot]]

Overview

This report summarises an automated exploration of the agentic-workflows agent across 4 representative software worker personas. 4 scenarios were tested; all produced high-quality, production-realistic workflow suggestions.

• Scenarios tested: 4 (Backend Engineer, DevOps Engineer, QA Tester, Product Manager)
• Average Quality Score: 4.9 / 5.0
• Run: §25565495554

Key Findings

• ✅ strict: true applied consistently across all 4 scenarios without prompting
• ✅ gh-proxy mode universally adopted — no scenario used local mode or raw MCP server startup
• ✅ Security posture maintained — agent job stayed read-only in every case; all writes routed through safe-outputs
• ✅ Bash allowlist scoping was context-aware — narrow lists on PR-triggered workflows (untrusted input), broad ["*"] on scheduled workflows (no untrusted input)
• ✅ Report routing worked correctly — scheduled + discussion-output scenarios were properly routed to the report prompt pattern (e.g., close-older-discussions, mentions: false)

Top Patterns

1. Most common triggers: pull_request (with optional paths filter) for reactive workflows; schedule with natural-language syntax (every Monday at 08:00 UTC, daily on weekdays) for periodic reports
2. Most recommended tools: gh-proxy mode + targeted toolsets (issues, actions, pull_requests) — never over-provisioned
3. Universal security practices: strict: true, read-only agent permissions, safe-outputs for every write, label/mention spam prevention (allowed-github-references: [], max-bot-mentions: 0)

View High Quality Responses (Top 3)

Backend Engineer — DB Migration Reviewer (5.0/5.0)

• Used paths filter on pull_request to avoid running on every PR
• Structured analysis with 🔴 Critical / 🟡 Warning / 🟢 Best Practice tiers
• Added duplicate-comment guard (checks for existing comment before posting)
• Correctly narrowed bash to [cat, grep, diff, find, wc]

QA Tester — Coverage Analysis (5.0/5.0)

• Artifact-first strategy avoids re-running tests unnecessarily
• hide-older-comments: true keeps PR clean across push events
• add-labels.allowed: [coverage-drop] prevents label injection abuse
• Three toolsets (default, actions, pull_requests) scoped precisely

Product Manager — Weekly Digest (5.0/5.0)

• mentions: false + allowed-github-references: [] correctly prevents notification spam from issue titles
• close-older-discussions: true + title-prefix prevents discussion accumulation
• Fallback behaviour for zero-issues weeks explicitly handled in prompt

View Areas for Improvement

• Cache-memory not surfaced: None of the 4 scenarios suggested using cache-memory for cross-run state persistence (e.g., storing a baseline coverage % between PR runs, or tracking failure trend data for the CI health report). The .github/aw/memory.md guidance could be more proactively recommended for stateful workflows.

• workflow_dispatch auto-addition undocumented in prompts: The compiler automatically adds workflow_dispatch to scheduled workflows, but this is not mentioned in agent responses. Users may be surprised when they see it in the compiled lock file — a brief note in prompt responses would improve transparency.

• Broad vs. narrow bash decision rationale: The DevOps scenario correctly used bash: ["*"] for a scheduled workflow with no untrusted input. However, this decision was made implicitly. Explicitly stating why (no user-controlled input in this trigger) would help users understand when to deviate from the default narrow allowlist.

Recommendations

1. Add a cache-memory usage example to .github/aw/memory.md targeted at PR workflows (e.g., storing baseline metrics between runs) and report workflows (e.g., tracking week-over-week trend deltas). Currently the agent does not proactively suggest persistence even when it would clearly improve workflow quality.

2. Document workflow_dispatch auto-injection in .github/aw/triggers.md so agents can mention it in responses — users should know their scheduled workflow can also be run on-demand without frontmatter changes.

3. Add a decision guide for bash allowlist scope to .github/aw/create-agentic-workflow.md — a simple rule: "narrow bash when the trigger exposes untrusted user input (issues, PRs, comments); bash: ['*'] is acceptable for scheduled or workflow_dispatch triggers with no untrusted input."

References:

• §25565495554

Generated by Agent Persona Explorer · ● 14.8M · ◷