Image: Entelligence AI
Why it mattersA team that runs a coding agent and then has to work out what it did spends real hours reading transcripts; a map and a timeline move that work to minutes.
A coding agent that fails leaves a long transcript behind, and the person who sent it on the task has to read all of it to work out what went wrong. Entelligence AI, which already sells code review tools for coding agents, released Agentplay this week to shortcut that reading. It reads the session logs on disk and draws what the agent touched as a map and a timeline.
Agentplay is a command-line tool that reads the session storage directories seven coding agents already write to on the machine that ran them: Claude Code, Codex in its CLI and IDE forms, Cursor's agent CLI and IDE composer, GitHub Copilot's CLI and VS Code modes, OpenCode and Kiro. It parses each session into events, redacts them, and serves a local web viewer that shows the agent's work as a radial tree of the files, tools, MCP servers, models and errors it touched, with a timeline below that plays the session back at one, four or sixteen times speed.
What the fidelity bar looks like
Entelligence says Claude Code sessions are the richest, because its JSONL log carries per-request token counts, hook calls, compactions, stop reasons and latency, and subagents get their own lane on the timeline. Codex is the next step down, with subagent splitting but no hook detail. Cursor, Copilot, OpenCode and Kiro record their tool calls but not their subagents, so those sessions render as one lane.
What leaves the machine, and what does not
Agentplay keeps raw prompts, tool inputs, tool outputs, file contents, command arguments, subagent prompts and user messages out of the map it writes. The map holds the shapes only: repository names, relative paths, tool and MCP names, model identifiers, token counts, timings and error classes. The viewer runs on 127.0.0.1 with no network calls, except an optional email signup that DO_NOT_TRACK=1 turns off.
The repository is Apache-2.0, created on 14 October, and holds about 1,000 Python tests across the parsers, the builder and end-to-end runs, plus more than 250 tests on the viewer. It is maintained by Entelligence AI, which also sells a hosted team edition that aggregates runs across a company.
Having a map of a run also makes a session discussable. A review question such as "why did the agent reach for grep four times before it wrote the fix" is a sentence that can be checked against a frame of the timeline, instead of a scroll through text. That is a cheap way to turn an opaque session into a specific one.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.