Officially launched

AI Agent Flow is now officially launched on Product Hunt.

aiagentflow CLI - A local-first AI engineering team in your terminal. | Product Hunt
Documentation

Agent Roles

Understanding all seven agents: Architect, Coder, Reviewer, Security, Tester, Fixer, and Judge.

AI Agent Flow uses a multi-agent pipeline to ensure high-quality, reliable output. Each agent has a specific role in the process β€” no single model handles everything.

Developer Insight: This multi-agent separation prevents "hallucination cascade" β€” where a single model might ignore its own errors. Each agent acts as an independent check on the previous one.

πŸ›οΈ The Architect

The Architect is the first agent to act. Its job is to:

  • Analyze the user request and relevant codebase context
  • Draft a step-by-step implementation plan
  • Identify potential edge cases and dependencies
  • Produce a detailed spec that the Coder follows

πŸ’» The Coder

The Coder follows the Architect's plan precisely. It:

  • Writes the actual code changes across one or more files
  • Ensures syntax correctness for the target language and framework
  • Implements the logic specified in the plan β€” no improvisation

πŸ” The Reviewer

The Reviewer acts as a second pair of eyes after code is generated. It looks for:

  • Logic bugs and off-by-one errors
  • Security vulnerabilities (injection, auth issues, unsafe defaults)
  • Style guide violations and anti-patterns
  • Missing edge case handling

If the Reviewer rejects the code, the Fixer agent is invoked automatically.

πŸ”’ The Security Agent

The Security Agent runs after the Reviewer approves the code and before tests are written. It scans the generated output for:

  • Injection flaws β€” SQL, command, LDAP, XSS, and template injection
  • Secret exposure β€” hardcoded credentials, API keys, or PII in logs
  • Broken access control β€” insecure direct object references, missing authorization checks
  • Cryptographic weaknesses β€” weak algorithms, predictable tokens, hardcoded salts
  • Dangerous function calls β€” eval, exec, shell passthrough, unsafe file operations
  • Security misconfiguration β€” open CORS, debug flags, default credentials

If the Security Agent returns FAIL, the Fixer agent is invoked with the findings as context β€” the same fix loop used for review rejections.

πŸ§ͺ The Tester

The Tester writes unit and integration tests for the generated code. It:

  • Generates test files covering happy paths and edge cases
  • Runs the project's configured test command (npm test, vitest run, etc.)
  • Reports pass/fail results back to the workflow engine

πŸ› The Fixer

The Fixer resolves issues surfaced by the Reviewer or Tester. It:

  • Reads the specific review feedback or test failure output
  • Makes targeted fixes without rewriting unrelated code
  • Re-triggers the Reviewer after each fix (up to maxIterations)

βš–οΈ The Judge

The Judge performs final QA validation. It:

  • Cross-references the output with the original user request
  • Checks that all Reviewer feedback has been addressed
  • Decides whether to PASS the task or send it back for another iteration
  • Only approves when the output fully satisfies the original requirements