How to Build a Custom AI Reviewer Agent (Performance, Accessibility, Style Cop)

Why custom reviewers

The default Agentopias by CynetIQ reviewer is a generalist — it will catch common bugs and style issues across any code. But teams almost always want specialist reviewers too:

• A Performance Reviewer that knows about N+1 queries, missing indexes, allocation hot paths.
• An Accessibility Reviewer that flags missing aria labels, contrast issues, keyboard traps.
• A Frontend State Reviewer that knows your Zustand patterns and flags useState misuse.
• A SQL Style Cop that enforces your in-house naming conventions on migrations.

Each persona is just a prompt + a model. Agentopias by CynetIQ gives you the harness to define them and pick which one runs on each task.

The five-minute setup

1. Create the agent

/dashboard/agents → New Agent. Set:

• Name: Performance Reviewer (or whatever).
• Role: pick one of reviewer, security_developer, qa, lead_developer — used for routing.
• is_reviewer: ON. (This makes the agent appear in the 🔎 Review picker.)
• Provider: OpenAI / Anthropic / Google / your self-hosted endpoint.
• Model: pick the right tradeoff. Performance review is mostly pattern-matching, so gpt-5-mini is plenty.
• System prompt: this is where the persona lives. See examples below.

2. Write the system prompt

Performance Reviewer (paste this):

``

You are a senior backend performance reviewer. For every patch, focus on:

N+1 query patterns (check ORM relationship loading).

Missing or unused indexes (look at WHERE/JOIN columns).

Synchronous I/O on the hot path.

Unbounded loops, list comprehensions, or in-memory accumulations.

JSON serialization of large objects.

Cache reads that fall through to slow backends.

Output format:

Summary

Findings (file/line, what's slow, magnitude estimate, fix)

Severity (critical/high/medium/low/clean)

Score (0-100)

Cite specific code. Do not comment on style or naming.

`

Accessibility Reviewer:

`

You are a WCAG 2.2 AA accessibility reviewer. Focus only on the diff. Check:

Interactive elements have keyboard handlers (button vs div onClick).

Form inputs have labels (aria-label or ).

Color contrast for new colors against panel-bg.

Focus traps in modals.

ARIA roles correctly applied (no role="button" on actual ).

Screen-reader-only context for icons.

Severity guidelines:

• critical: keyboard inaccessible critical flow
• high: missing label / WCAG fail on auth/form path
• medium: WCAG fail on non-critical UI
• low: nit

`

3. Pick the agent on a task

Click 🔎 Review on any task. The picker shows:

• ✨ Auto (resolves from task tags / Integration Rules)
• Default Reviewer
• Performance Reviewer
• Security Developer
• Accessibility Reviewer
• (any other is_reviewer agent you've defined)

Pick yours. The review runs synchronously and the report appears on the task's Reviews tab.

4. (Optional) Auto-route

For a persona to fire automatically on certain tasks, define an Integration Rule:

• "All Sentry issues from the backend-api project → Performance Reviewer"
• "All Jira tickets with label accessibility → Accessibility Reviewer"
• "All Azure DevOps work items in Area\\Frontend → Frontend State Reviewer"

Same rule engine, every source.

Audit history

Every reviewer agent has its own history banner on the agent card: severity distribution, average score, click-through to past reports. Filterable on /dashboard/reviews.

Pro tips

• Run a cheap model for nits, an expensive model for security. gpt-5-mini for the SQL style cop, gpt-5-pro` for the security_developer.
• Iterate on prompts in Prompt Studio. Roll back to previous versions if a prompt change degrades quality.
• Don't mix concerns. A reviewer that checks performance AND security AND style will be mediocre at all three. Build narrow personas.

Related reading

• AI Code Review landing page
• OWASP-aware AI code review tutorial
• From Sentry alert to merged PR in 12 minutes