Coding sessions
The step that repeats in every session.
Sessions' goals never repeat — but the exploration, triage and survey steps inside them do. Those are the ones worth learning.
Anatomy of the task · illustrative workflow
Locate files, symbols and call sites; report paths, not dumps.
Input
A codebase, a question about it, and the read-only tools an exploration step is allowed to use.
Required output
The relevant paths and symbols, reported precisely — not a dump of file contents into the context window.
Evaluation
- Names the right files and symbols
- Reports paths, not dumps
- Leaves the session cheaper than a frontier pass would
Understand the workflow.
On our own development estate, an Explore-style sub-agent thread recurred in 71% of coding sessions. The sessions' goals never repeat; the sub-agent's does. That is the shape Operant looks for.
Define the quality bar.
An exploration step has a checkable outcome: did it find the right call sites, and did the parent session succeed downstream of it? Score that, on held-out sessions, against what the frontier model found.
Improve the route.
The recurring step becomes a named pattern. Operant writes the skill it takes — which tools, what to report, what to skip — and proposes routing that step to a cheaper model carrying it.
Keep frontier capability.
The main thread keeps the frontier model. Only the recurring sub-task moves down, and the frontier model is always the failover when the cheaper route stumbles.
Published evidence
See the comparison. Keep the context.
The published gate result shows the discipline: exemplars held out before the learner ran, disjoint ids, and a judge scoring parity against the frontier baseline.
Read the results ->Who this fits
A task your team knows well.
A team running Claude Code, DeepSeek Harness or Codex sessions at volume, where the same exploration and triage steps recur across sessions.
Bring a workload ->Another place to start · Scheduled digests