Baz Agents

Coding and Code Review agents that analyze pull requests for quality, correctness and security issues.

Baz runs AI agents, including coding agents and code review agents, as for evaluating pull requests. Rather than assessing changes on a per-file basis, these agents consider the entire repository and its external context. The codebase is split into indexable units, and embeddings with similarity measures are used to retrieve relevant code and tests. Agents perform agentic code analysis and optional runtime inspection, yielding structured findings. These findings are shared as pull request comments and CI check results.

Coding Agents

Fixer is a new class of code composing agents. It runs in a ephemeral, secure, sandboxed environment and can be tasked with fixing issues discovered during review.

Accelerates the code review cycle by letting suggested fixes be committed directly to a PR, eliminating manual edits and context switching.

What it does

  • Proposes small, safe edits to address verified issues and can apply them to the PR so reviewers and authors see a working suggestion in-place.

  • Focuses on self-contained fixes that are low risk to apply automatically.

High level guidance

  • Only apply fixes that are clearly correct and scoped to the change. Avoid risky changes that require design or product decisions.

  • Keep the suggested changes minimal and accompanied by a short rationale so reviewers can accept or tweak the suggestion quickly.

Tools it uses Tools that gather code and diff context, tools that produce a patch or suggestion, and tools that safely create commit suggestions against the PR.

Context it consumes PR diff, related files needed to justify the fix, and any metadata that explains intent (PR title or ticket). The agent favors fixes that can be validated by the changed code alone.

How it behaves Runs a quick verification workflow: build a minimal justification for the change, prepare a patch, and surface the patch as a suggested commit. The agent favors simplicity and high confidence fixes.

How to Trigger Once Baz Fixer is enabled and configured, you can trigger fixes directly from your PR in two ways:

  • Apply fix on a single comment Each Baz review comment includes a checkbox: “Apply fix with Baz”. Selecting it will generate a commit that addresses that specific finding.

  • Fix all comments in a PR

    If your PR contains multiple Baz comments, you will see a “Fix all” option in the PR description. Selecting it will generate a separate commit for each open Baz comment in the PR.

Code Review Agents

Reviews are our general purpose code-review agent class. They are individually scoped, contextualized and steered to discover, analyze and fix coding issues on specific engineering sub-domains. Combined with memories, derived by user feedback to the Baz agent on pull requests, each agent is both extremely focused and highly tuned to your codebase's unique requirements.

Ensures implemented code and design align with documented requirements, identifying gaps or deviations early.

What it does

  • Extracts explicit requirements from tickets and designs and validates whether the implementation satisfies those requirements.

  • Produces a verdict for each requirement with evidence: met, partially met, or not met.

High level guidance

  • Keep extraction strictly ticket-driven: only record requirements explicitly stated in the source materials.

  • Validate each requirement using code and, when available, preview environments or design artifacts.

Tools it uses Tools that fetch ticket and design artifacts, tools that help get context from code and diffs, visual comparison helpers when preview environments are available, and evidence capture tools.

Context it consumes Ticket text and attachments, design files, PR diff, optional preview environment snapshots, and prior specifications for consistency.

Activation note Connect your integrations to activate this agent. When design or preview integrations are present the agent will include visual validation as part of the verdict.

Code Correctness

Identifies logical inconsistencies, flawed conditionals, and edge cases that could produce unexpected behavior.

What it does

  • Highlights incorrect logic, incomplete implementations, missing steps, and unintended side effects.

  • Gives concrete traces and examples of failing execution paths.

High level guidance

  • Compare the implementation with PR intent or ticket context to determine whether behavior is intentional. Prioritize concrete, reproducible issues.

Tools it uses Tools that map code flows and help extract execution traces along with code and diff exploration utilities.

Context it consumes PR title and ticket context, diff hunks, and the code paths needed to trace complete execution from input through output.

Code Quality and Correctness

Finds unclear identifiers and obvious spelling mistakes that reduce code clarity.

What it does

  • Flags non descriptive or incorrect names and typos in code and comments.

High level guidance

  • Be conservative with stylistic nitpicks. Avoid enforcing strict naming conventions that conflict with the repo style.

Security Best Practices

Identifies common security anti patterns like unsanitized inputs, PII exposure, and injection vectors.

What it does

  • Flags hardcoded secrets, PII leaks in logs, risky SQL or command usage, and missing input validation.

High level guidance

  • When calling out PII or secrets, specify exact locations and avoid hedging language. Avoid flagging framework handled behavior or test placeholders.

FAQ

chevron-rightWhat do Baz’s default agents do?hashtag

They analyze change requests for naming, typing, logic bugs, outdated comments, log errors, etc., using a combination of AI, parsing, and repository context.

chevron-rightHow does Baz scale efficiently on large codebases?hashtag

Baz divides code into manageable chunks, reprocesses only changed files, stores embeddings in a vector database for similarity search, and filters by organization/repo to maintain performance.

chevron-rightCan I disable some default reviewer checks?hashtag

Yes. Organization admins can deactivate specific agents or modify their scope.

Last updated