Entire has launched Agentic Search, a search system for code and engineering context. It is designed to let software agents search across every repository they can access.
The system combines exact code search with semantic search across sessions, commits, checkpoints and transcripts, and exposes those functions through a single API. Results return matching code and commit data, along with the session, transcript and prompt linked to a change.
Founded by former GitHub Chief Executive Officer Nat Friedman, Entire is positioning the tool around a problem that has become more visible as software teams use AI agents in development work. Traditional code search tends to focus on symbols, strings and files. Newer systems also need to trace how a package was used across an organisation, which approach a team adopted and why a change was made.
That information often sits across repositories, commit histories and records from earlier coding sessions. In practice, developers and agents may need to clone repositories, search files manually and review commits one by one to reconstruct a decision.
Two search modes
Agentic Search was built to serve two tasks. One is semantic search, intended to answer questions about what happened and why by scanning sessions, commits, checkpoints and transcripts for meaning rather than exact wording.
The other is literal code search, which looks for exact lines of code and is described as symbol-aware, with definitions and references. Agents can use both forms of search and either limit queries to individual repositories or search across all repositories available to them.
As an example, Entire said a search for "rate limit retry logic" could find the session in which an agent wrote the logic even if the word "retry" did not appear in the code itself. That reflects a broader attempt to connect generated software output with the workflow that produced it.
Internal benchmark
To support the launch, Entire published results from an internal benchmark based on nine engineering-history questions drawn from its own organisation. The questions focused on what the company described as one of the hardest problems for agents to answer: why code was written in a particular way.
In the test, two otherwise identical headless Claude agents were run 10 times on each question. Both had access to full-history clones of the relevant repositories, Git log and show commands, string search, and the GitHub command-line interface for pull requests and issues. One agent also had access to Agentic Search.
Across 90 runs, the agent using conventional repository-history tools answered 70 correctly, while the agent with search answered 81 correctly. Entire said the search-enabled version also used an average of 262,000 tokens per question, compared with 547,000 for the baseline. Mean cost per question fell to USD $0.23 from USD $0.38, and average agent steps dropped to seven from 14.
One highlighted result concerned what Entire described as the deepest-history question in the set, focused on the design history of an autoscaler. On that task, the Git-history-only agent scored 0 out of 10, while the agent with search scored 10 out of 10. Median response time was 35 seconds with search, compared with 407 seconds without it.
Entire said the search-enabled agent used the tool in all 90 runs without being explicitly instructed to do so, and made it the first tool call in 66 of them. The company argued that this suggests agents can identify when the answer to a "why" question is more likely to be found in sessions and checkpoints than in source code alone.
Caveats noted
The benchmark also came with limits. Entire said the questions were taken from its own engineering history, creating the possibility of task-selection bias. It also noted that its commit messages are relatively detailed because agents write many of them, which may have improved the performance of the non-search baseline.
That means the comparison may not translate directly to other software teams, although Entire said the gap could be wider in environments with less structured history. It also said it had published the benchmark harness, tasks, grading method and per-run results so other developers could rerun the test in their own organisations.
The launch adds to a growing push among developer-tool suppliers to make AI systems work with more than static codebases. As companies test autonomous or semi-autonomous software agents, the challenge is shifting from code generation alone to retrieving reliable context about earlier design decisions, implementation patterns and operational history.
Entire's pitch is that agents should not just inspect the final state of a repository but also query the work that led to it. As the company put it, each result returns "the matching code and commit, plus that commit's checkpoint: the session, the transcript, and the prompt behind the change."