On 8 November, InfoQ journalist Sergio De Simone reported on Entire’s Git service, which is built for collaboration between people and coding agents. The company’s starting point is that the pull request model does not fit a world in which agents write, review and fix code at a fast pace.

Entire’s CEO, Thomas Dohmke, argues that GitHub and GitLab built human collaboration around asynchronous code review, InfoQ reported. With agents, however, the focus shifts to steering, versioning and auditing their work.

Scores appear before the merge

At the core of the service is the trail, a workspace for a change that can span several repositories and be started from a bug report, a feature request or by hand. InfoQ quoted Entire: ‘The pull request [timeline] has become an audit log that only the agent reads.’

Whenever new code is pushed, a review agent running in a cloud sandbox gives the change a score with a rationale. It assesses code quality, security, slop (carelessly machine-generated code) and whether human review is needed. Each evaluation shows the five previous runs alongside the current one, because a single pass often misses a critical problem, Entire said.

All findings must be resolved or closed before a merge. Approval conditions can be set per repository. The service’s distributed architecture makes pushes up to 89 times faster than pushes to GitHub, InfoQ reported. InfoQ does not say how the figure was measured.

Memory from one agent session to the next

Entire says persistent memory keeps accumulated knowledge in the repository so agents can use it across sessions. The open-source tools entire-graph and entire-brain provide this memory. InfoQ reported that entire-brain links a code map to earlier agent sessions and Git history.

The entire-graph GitHub page says the tool parses a repository locally without network connections or model calls. It gives the agent a map of definitions, callers, types and the impact of a change, with line numbers. Semantic parsing covers 36 languages.

The company says it measured a score of 94.74 per cent on the LoCoMo benchmark, against 93.83 for mem0, 92.86 for cognee and 91.88 for a bm25 baseline. LoCoMo measures whether a tool finds the right passage in long conversations, not in code. The test used an unmerged development branch of entire-graph.

The company also says token savings reach up to 71 per cent, depending on the situation. In its own paired comparisons, however, token use rose when the agent ran the graph search as a tool call. When the graph search was added to the prompt in advance, the difference was +0.7 per cent and not statistically significant.

The page warns that static analysis is heuristic. Calls that pass through interfaces or dynamic binding may go undetected. In an example run, the tool reported zero callers for a function even though the source code contained two.

The MIT-licensed entire-graph is an extension of the Entire CLI and requires at least CLI version 0.10.0 and Git version 2.36. Users enable it one repository at a time: the tool writes its agent instructions to a separate file and adds a marked block to AGENTS.md and CLAUDE.md without changing the rest of the text.