Simulation
Several language models hold moderated conversations, with a chair, a watchdog and a tool broker.
The simulation lets several agents talk to each other: moderated or round-robin, with a chair, a watchdog and a broker for tools. It is provider-agnostic and runs on cloud models as well as local ones.
In production since 18 May 2026.
What it is
Several agents, one table
A dialogue engine: several language models take turns on a question, moderated or round-robin, with three fixed roles at the table: a chair, a watchdog, a tool broker. The chair keeps the thread, the watchdog stops drift, the broker hands out tools when an agent asks for one.
Any provider
Anthropic, OpenAI-compatible, OpenRouter, local (Ollama). The same run can seat a cloud model next to a local one; what matters is the transcript, not the vendor.
What it is for
Trying a question from several sides before anyone commits to an answer: a design decision, a reading of a situation, a claim from the findings engine. The transcript is kept with the run, so the argument can be re-read later.
Since
In production since 2026-05-18. Transcripts are not shown here: the runs so far argued about the lab's own work, and that stays inside.
What has run
Counted 3 September 202632 real runs across 9 setups between 18.05.2026 and 31.07.2026, read from the run logs on the agent's host. Beside them, 90 daily smoke runs against a local model, which only check that the engine still works end to end. The largest real run: 5 agents, 19 turns, 608,483 tokens, on 20.05.2026.
| how the real runs ended | runs |
|---|---|
| completed | 11 |
| hit the turn budget | 8 |
| stopped by the watchdog | 7 |
| stopped by the operator | 5 |
| engine crash | 1 |
Every ending is a row, including the ones that were not planned. A watchdog that stops a run is doing its job. The engine crash stays in the count; the engine has since been built to survive one without taking the parallel runs down.