Atlassian shipped Code Context, an AI context engine for the whole codebase, and the problem statement is the same one every agent platform has arrived at this month from a different direction: “when an agent starts in the wrong place, everything slows down and costs more.” Their answer is to index Bitbucket and GitHub repositories into the Teamwork Graph with both lexical and semantic search, then wire the code to the signals already sitting around it — Jira issues, Confluence pages, Loom recordings, Slack threads, and 50-plus third-party connectors. The claim is 44% more accurate results using 48% fewer tokens, which is a vendor benchmark on a vendor’s own graph and should be read as directional. The architectural decision underneath it needs no benchmark to be interesting.
That decision is permissions. Indexing respects existing SCM permissions, and “results are scoped to what each user or agent can already access” — the agent inherits the caller’s reach rather than being handed a service account with a view of everything. Indexing is admin opt-in, not on by default. And they draw the accountability line explicitly rather than letting it blur: “every code change still needs an owner… developers are accountable for understanding the change, reviewing downstream impact, and deciding what ships,” because “the goal is not to give agents unlimited access.” For a company whose entire commercial incentive points at maximum agent autonomy, publishing a ceiling on it is the notable move.
The catalog holds 183 API pages for Atlassian, which is what an estate looks like when the product is other people’s workflows, and Code Context is stitching across the seams between them. The Bitbucket Source API and Bitbucket Commits API are the code being indexed; the Jira Software Development Information API is the existing join between a commit and the issue that asked for it; the Jira Software Epic API and Sprint API hold the business rationale the post says agents lack; and the Bitbucket Branch Restrictions API is where “developers stay in control” stops being a sentiment and becomes a rule.
Atlassian scores 65.6, strong on the Kin Score — contract quality 72.4, developer ergonomics 70.2 — and Agent Readiness is 47.3, agent-ready on eleven lit dimensions. Three of them corroborate this post directly rather than sitting beside it: delegated_identity is lit, which is the machine-readable form of “scoped to what each user or agent can already access”; dry_run_mode is lit; reversibility_documented is lit. A company claiming that agents inherit permissions and changes stay reversible has both of those legible in its contracts, which is rarer than it should be. The gap is narrower and specific. event_surface_described is unlit — on a platform whose Teamwork Graph is built from signals other systems emit. A graph fed by events, whose event surface is the thing an outside integrator cannot read.