Atlas Upstream
Follow an agent from the first request to the evidence that settles it. Each step replays an actual decision from a recorded run of the prototype.
Version 0.2 adds seven discoverable MCP tools, atomic evidence batches, restart recovery and an Opsbook lesson export. Its new protocol trial wrote a real file, detected stale contents, repaired them and recovered the verified result in a fresh server process.
Install and connect an agent
Extract the download, open the atlas-upstream folder and install with Python 3.10 or newer.
python3 -m pip install ".[mcp]" atlas-upstream-mcp --workspace /absolute/path/to/project/.upstream
Configure an MCP host to launch that command. It discovers seven tools for creating missions, recording evidence, resuming work and exporting lessons. Read the host setup guide. A Python wheel is also available.
44 tests passed, including a real file-repair task through the official MCP client. Live language-model trials and physical field trials remain pending. Explore Atlas Opsbook for testing and retaining proposed lessons.
Journal entries since the previous step
What did fewer questions buy?
Sixteen simulated handovers varied the destination, access requirement, host’s preference and lift availability. Both policies used the same route planner. One asked when an answer could affect a decision. The other collected every input before setting off.
A human contact is a modelled interruption. Friction combines declared time, effort, interruptions and applicable preference penalties. The first physical trial with people is still to be run.
| Across 16 cases | Selective | Eager |
|---|---|---|
| Verified handovers | 16 | 16 |
| Information requests | 40 | 80 |
| Modelled human contacts | 32 | 64 |
| Weighted friction points | 324 | 493 |
| Hard-constraint violations | 0 | 0 |