Entire has released a search system designed to help AI software agents find their way through years of repository history, commit logs, and the transcripts of earlier coding sessions.
How Agentic Search Works
Agentic Search combines two approaches. One mode scans for exact code patterns—definitions, references, symbols—across every repository an agent can access. The other mode searches for meaning: it can surface the session where an agent wrote retry logic for rate limiting even if the word “retry” never appears in the code itself. Both modes feed into a single API that returns matching code, the associated commit, and the checkpoint—session, transcript, and prompt—behind that change.
The company was founded by Nat Friedman, who previously served as GitHub‘s chief executive. Entire is positioning the product around a gap that has become more apparent as teams deploy AI agents in development work.
The Problem with Traditional Search
Traditional code search handles symbols, strings and files adequately. What it does not handle well is tracing why a team chose a particular approach, how a library came to be used across an organization, or what decision led to a given piece of code. That information tends to live across repositories, commit histories and earlier coding session records. In practice, developers and agents often resort to cloning repositories, searching files manually and reviewing commits one by one.
Internal Benchmark Results
To test the system, Entire ran an internal benchmark built around nine engineering-history questions drawn from its own organization. The questions focused on what the company described as one of the hardest problems for agents: understanding why code was written in a particular way.
Two headless Claude agents—identical except for tool access—ran each question ten times. Both had access to full-history repository clones, git log and show commands, string search and the GitHub command-line interface. One also had access to Agentic Search.
Across 90 runs, the baseline agent answered 70 questions correctly. The search-enabled agent answered 81 correctly. The search version also consumed an average of 262,000 tokens per question, compared with 547,000 for the baseline. Mean cost per question dropped to USD $0.23 from USD $0.38, and average agent steps fell to seven from 14.
One result stood out. The deepest-history question in the set concerned the design history of an autoscaler. The agent using only Git history scored zero out of ten on that task. The search-enabled agent scored ten out of ten. 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 those runs. The company argued this suggests agents can recognize when a “why” question is more likely to be answered in sessions and checkpoints than in source code.
Benchmark Limitations
The benchmark carries limitations. The questions came from Entire’s own engineering history, which introduces the possibility of task-selection bias. The company also noted that its commit messages tend to be detailed because agents write many of them—a factor that may have helped the baseline perform better than it would in environments with less structured history. Entire has published the benchmark harness, tasks, grading method and per-run results so other developers can run the test in their own organizations.
Market Context
The launch fits into a broader shift among developer-tool suppliers. As companies test autonomous or semi-autonomous software agents, the challenge is no longer just generating code. It is retrieving reliable context about earlier design decisions, implementation patterns and operational history. Entire’s pitch is that agents should be able to query not just what a repository contains, but the work that produced it.