We are entering an era where agents will be writing, reviewing, and debugging most of the code. That changes what a well-maintained codebase looks like.
Anthropic's engineering team describes progressive disclosure as a core pattern for effective agent context: agents should "incrementally discover relevant context through exploration" rather than loading everything upfront. The goal is finding "the smallest set of high-signal tokens that maximize the likelihood of your desired outcome."
PR descriptions, commit messages, and review comments are the layers of progressive disclosure in your codebase. When they're empty, agents skip straight to expensive full-context exploration. When they're structured, agents work faster and make better decisions.
At Gitar, we like to think about it as giving the agent:
- Eyes: structured context like PR summaries, commit messages, and review history that lets agents understand what's happening and why
- Hands: the ability to act on what they learn, from applying fixes to encoding team patterns into reusable workflows
- Legs: access to systems beyond the code, like issue trackers and CI, so they can find context that the repo alone doesn't have
Let agents see what you see
Tribal knowledge hurts agents the same way it hurts new team members. When the reasoning behind a change lives in someone's head instead of in the PR description, everyone else has to guess. Agents guess by exploring a lot of files. PR descriptions end up as commit messages and live directly in git, so a structured summary there lets anyone understand a change without that exploration.
# Agent-unoptimized
Title: Refactor authentication flow
Description: "This PR cleans up the auth module and improves
error handling. Should make things more maintainable going forward."# Agent-ready
Title: Refactor auth flow to use centralized error handling
## What changed
- Extracted error handling from AuthService into ErrorHandler
- Replaced 7 scattered try-catch blocks with error boundaries
- Added typed errors (AuthenticationError, TokenExpiredError)
## Why
- Inconsistent error responses (some 500s, some 401s, same root cause)
## Breaking changes
- PaymentService.charge() now returns Result<Receipt, PaymentError>
## Related
- PROJ-1847 (original ticket)
- PR #11890 (initial discussion)
Scan the title, read the summary, drill into specifics only if needed. Progressive disclosure.
At Gitar, we enhance descriptions in-place without losing the original. We pull in linked tickets and related PRs so summaries are shaped by more than just the code diff. When the existing description is already comprehensive, we leave it alone. Context should live where it gets read, not in a collapsible comment that gets skipped.
Summaries are one example. The same applies to test descriptions, module-level doc comments, cross-PR links, commit message formats. Anywhere you can add structured context is a place where you're giving agents better eyes.
Let agents act on what they learn
Code reviews are where a lot of institutional knowledge gets created. A quick "LGTM" gets the PR merged, but a review that says "This O(n^2) loop will be slow past 1K items, we hit this in PR #847, consider a HashSet" is something anyone can search for and learn from later.
Reviews that explain why something was approved or rejected capture reasoning that doesn't belong in the code itself but is essential for understanding it.
At Gitar, our agent learns from past reviews, feedback, and resolution rates to understand what patterns work in a codebase and what doesn't. It applies fixes, enforces standards like logging consistency and naming conventions, and opens PRs against your agent skills and workflows. The richer the review history, the better it gets at matching the team's expectations.
Locally, tools like Claude Code support skills, which are reusable workflow files that encode team knowledge into steps an agent can follow. A skill for running a deploy, debugging a flaky test, or writing a migration lets your team share learnings once and re-run them without repeating steps. Same idea: captured knowledge that agents can pick up and act on.
Let agents go find what they need
Context doesn't always live in the code. Consider a PR description that just says "Addresses PROJ-1847." An agent can see the mechanics of what changed in the diff, but without access to the ticket, it can't understand the business motivation or the customer issue that triggered the work. Making sure the code does what the business needs it to do requires understanding intent, not just implementation.
This is where legs come in. Ticket trackers, telemetry, crash reports, CI logs: a lot of the signal agents need is spread across systems. At Gitar, we close these loops across the workflow. We auto-link PRs to relevant issues even when developers forget to tag them, add implementation summaries back to linked tickets when changes merge, create follow-up issues for unresolved comments and TODOs so nothing gets lost after merge, and retry flaky CI jobs so developers don't waste cycles on intermittent failures.
When agents can connect to your integrations and pull context beyond the code, they make better decisions.
It's easy to get started
These standards have always existed and helped humans maintain codebases. With agents writing more code every day, it's more important than ever to make your codebase agent ready.
The good news is you don't have to do all of this manually. Install Gitar and you get an agent with eyes, hands, and legs: enhanced PR summaries, reviews that learn from your team's patterns, and integrations that keep your issues, CI, and code in sync. Your codebase becomes agent-ready with every PR.