Why use an AI agent to read code?

Because grep tells you where a symbol is, not how it behaves end to end. An agent that has indexed the repo can trace a request from route to database and summarise the flow. For a new codebase, that turns a day of reading into an hour of questions.

How do you ask good questions?

1. Start broad: “What does this service do and what are its entry points?”
2. Narrow to flow: “Trace what happens when a user resets a password.”
3. Ask for files: “Which files would I change to add a new field to the user model?”
4. Verify against the code before acting — cross-check the files it names.

Treat answers as a map, not the territory. The agent can miss dynamic imports and config-driven behaviour, so open the files it points to.

How does this change onboarding?

• New hires get answers before interrupting a teammate.
• Reviewers check a claim in the PR against the agent before arguing about it.
• Maintainers audit dead code by asking what still calls a module.

The habit that pays off is asking before editing, not after — it catches wrong assumptions early.

How reliable are the answers?

Good on structure and local logic; weaker on anything that depends on runtime config, feature flags or external services it cannot see. If a question hinges on production behaviour, the agent gives a hypothesis and you still need a log or a test to confirm it. Never merge a risky change on the strength of an agent answer alone.

When is it the wrong tool?

Skip it for repos with heavy generated code or monorepo sprawl where indexing is partial, and for legal or licensing questions it cannot judge. It also cannot replace reading critical security paths yourself. Use it to navigate, then verify.

Quick Summary

• Ellipsis also works as an agent that answers questions about your repo.
• Ask broad, then trace flows, then request the exact files to change.
• It speeds onboarding and lets reviewers verify claims before debating them.
• Answers are a map: strong on structure, weak on runtime and external behaviour.
• Skip it on heavily generated monorepos and never merge risky changes on its word alone.

What does Ellipsis review?

It reads the pull-request diff plus surrounding code and posts review comments: style issues, potential bugs, missing tests, and logic that contradicts earlier patterns in the repo. It also acts as a chat agent you can ask “where is auth handled?” and get an answer grounded in your code.

It fits small teams without a dedicated reviewer on every PR. It does not replace human review on security-sensitive paths.

How do you install it on a repo?

1. Sign in with GitHub and authorise the Ellipsis app.
2. Choose which repositories to enable. Start with one active repo, not all of them.
3. Open a test pull request and watch the first review comment appear.
4. Reply to a comment to check the conversational loop works.

Enabling every repo at once floods you with comments before you have tuned the rules. One repo first.

How do you tune the noise level?

• Add repo rules describing your conventions, so it stops flagging things your team does on purpose.
• Tell it to ignore generated files and lockfiles.
• Mark comments as unhelpful; feedback shapes later reviews.

Expect a week of tuning. Out of the box it errs toward commenting; after tuning, a good review should have a handful of substantive notes, not thirty nitpicks.

Is Ellipsis free?

Ellipsis has a free tier for small teams or limited usage, with paid plans for higher review volume, private repos at scale and team features. Check current limits, since review-based pricing changes as models improve. Budget by pull-request volume, which is what actually drives cost.

Where should you not trust it?

Anything security-critical, licensing-sensitive, or architecture-level. AI review is good at local bugs and missed cases; it is weak at judging whether a design decision is right for your product. Treat its comments as a first pass, and keep a human as the approver.

Quick Summary

• Ellipsis reviews PRs and answers questions about your codebase.
• Install on one repo first, then expand after tuning rules.
• Reduce noise with repo conventions and by ignoring generated files.
• Free tier covers small usage; paid scales with review volume.
• Do not rely on it for security, licensing or architecture calls — keep a human approver.