An agent work graph: connect AI coding history to the work it produced
Chat logs, browser research, file changes, and reviews tell different parts of one task. Here is a way to connect them without confusing a recorded event with a proven decision.
この記事を日本語で読む →
One task leaves traces in many places
A developer may investigate an issue in a browser, ask Claude Code for an implementation, have Codex review it, adjust the code in Cursor, and open a pull request. Each tool has a history. None necessarily explains the whole task on its own.
An agent work graph is a way to relate those traces: the question, evidence, agent conversation, attempted change, test, review, and result. The term describes a useful model for work history; it does not mean that every event has already been connected automatically or that a graph can prove why a person made a decision.
Record events, then connect them carefully
| Trace | What it can establish | What still needs checking |
|---|---|---|
| Agent conversation | What was requested and discussed | Whether the suggested change was applied |
| Tool or terminal output | Which command ran and what it reported | Whether the environment and result are still current |
| File diff and commit | What changed in the repository | Whether the change came from the cited conversation |
| Browser and app activity | Which pages and windows were open | Whether the page was read, trusted, or used in a decision |
| Review and test | What was checked at that point | Whether later edits invalidated the result |
The connections are more useful than a larger transcript. A link between an agent suggestion and a commit should be supported by the actual diff. A visited URL is a research lead, not proof of agreement with the page.
The useful unit is a piece of work
Ask: *What changed on this issue, what evidence informed it, and what remains open?* A useful answer groups events by task and time. It names the relevant file or pull request, separates completed checks from unverified steps, and points back to source material.
This is especially helpful when work crosses agents. The next agent can start with a short handoff, then inspect live files and tests instead of treating an earlier summary as current truth.
How Contextberg fits
Contextberg records app and screen activity, browser history, typed input, and supported agent conversations locally when those capture options are enabled. Record gives you the underlying timeline. Remember, Memory, and Chat use it to help reconstruct work and propose a place to resume. Its MCP connection can make selected context available to compatible agents.
The graph remains an aspiration about how to relate work, not a claim that Contextberg automatically attributes every code diff or GitHub artifact today. Start with a checkable timeline and add relationships only when the evidence supports them.