Working with agents

The papers, the crucible experiments, the FMT wiki and this website are all worked on in Claude Code sessions run through Agent Fleet. Agent Fleet is an open-source configuration layer by the same author, used every day across 8 machines and about 78 projects. What an agent session works out is lost when the session ends unless it is written down. Three habits address this, and each bears on whether the work can be checked afterwards.

Written handovers

Each session keeps a record of its state, its decisions and its open work. At shutdown the record is archived to a session log, and decisions go into a record of their own. The next session starts from the written handover. Anyone can later read what was decided and why.

Independent re-checks

Projects are kept apart. A research session cannot write into the infrastructure project, or the reverse, so work crosses between them as messages. Because of this separation, a second session with its own context can check what the first one did.

In October 2026 the research and infrastructure sessions launched a survey study together. The infrastructure session imported it, smoke-tested it and activated it. The research session re-checked independently and found that the test cleanup had deleted more rows than the test had created. The fix was applied and acknowledged within the hour.

Tests before code

crucible commits hypotheses and analyses before the data exist, and counts null results. The tooling follows the same order: code is written against a failing test, and the open-source Agent Fleet repository carries about 3,000 automated tests.

Details

The design, the installation steps and the safety rails are described on the main site: matthiasgruber.com/agent-fleet/. Code: github.com/JeltzProstetnic/agent-fleet, MIT licence. Questions: research@matthiasgruber.com