
Summoned for a decision, it reasons and emits the intent — plans and swarm specs — that the rest of the squad executes. View agent →
The local AI agents we run and study. Each is evaluated through the same fixed harness.

Summoned for a decision, it reasons and emits the intent — plans and swarm specs — that the rest of the squad executes. View agent →

Given an issue and the relevant source, produces an implementation specification that the executor and verifier agents act on. View agent →

Takes a single scoped step and writes the code for it — the only agent with file-write access, gated by an unconditional compile check. View agent →

Reads an implementation and the test that claims to prove it, and returns a verdict — built to reject hollow tests. View agent →

The read-only researcher, fanned out in parallel to gather from many sources at once — cheap, wide local search that feeds the judgment step. View agent →

The deep-research ingest wizard. Point it at a corpus or a question and it reads everything, condenses the signal, and returns a synthesis ready for judgment. View agent →
The 3090 node's standing researcher — dispatched for long, patient research turns and for triage when the lab's own checks flag something worth a closer read. View agent →

The always-on caretaker, resident on the machine that hosts the substrate itself — it extracts and consolidates the day's memory, watches RAM and serving budget, and reports squad health, for almost nothing. View agent →

The chronicler. Reads each hour's transcripts, distils them to what is worth keeping, and writes the record to the substrate so the work outlives the session. View agent →