Receives a snapshot of the experimental state (e.g., leaderboard, experiment history, other subagents' work)

Returns a decision.md for other agents to execute on.

• Reads the failing eval diagnosis to identify failure modes and missing capabilities.
• Searches relevant research, and specs out the next experiments with a hypothesis, expected effect, blueprints for synthetic data generation, and search tasks for the scout.
• Every time the performance plateaus, the loop agent spawns a new Strategist with fresh context, and all available experimental memory.

Receives a data spec (e.g. "olympiad math questions with verified reasoning traces")

Returns training assets.

• Searches HuggingFace and the open web, reporting a shortlist of candidate datasets.
• Onboards the selected candidates into the data catalog and decontaminates each against all eval sets.
• Registers every asset with lineage, so any experiment's data traces back to its source.

Receives directional guidance for synthetic data generation, such as seed data, target capability, and verification method.

Returns verified synthetic datasets.

• Builds the synthetic pipeline infrastructure: preprocessing, prompt building, generation, and answer verification.
• Generates with a teacher model via external API or serves a previous checkpoint to self-distill.
• Iterates in small batches to refine quality before scaling up quantity.

Receives an experiment's eval results on a benchmark.

Returns per-case failure analysis.

• Multiple diagnosis agents are spawned in parallel, each focussing on one benchmark.
• Reads each example and records generalizable failure modes (e.g. arithmetic errors, context length exceeded, broken reasoning).
• Re-runs failing code against the unit tests to find specific error in model outputs.