How DistilCode saves tokens
DistilCodeBack to homeHow they are saved
Some optimisations used:AutoJev model selection
The first normal prompt in a new session is routed to a model class suited to the work.Explore techniquePrompt and context caching
Stable prompts and shared context can be reused instead of processed from scratch on every call.Explore techniqueLossless context pruning
Old conversation is compacted while actionable facts, constraints, and file state are carried forward.Explore techniqueDynamic delegation of tasks
Bounded work can move into a specialist child session with an appropriate model class and tool permissions.Explore techniqueTool verbose reduction
Oversized tool output becomes a bounded preview while the complete result remains available for targeted inspection.Explore techniqueCodebase graphing and traversal
A repository index gives exploration agents structural lookups before they resort to broad file reads.Explore techniqueTask-specific tool adaptation
Tool definitions and permissions adapt to the active model, client, feature flags, and specialist role.Explore technique
How they were evaluated
Built to mirror real work.
SWE-bench
Evaluated on SWE-bench—real GitHub issues, real repositories.700+ runs
Repeated for redundancy (the developer went broke testing the framework).Real agentic work
Models solved issues the way they would in practice—in a live repository.Same LLMs, two harnesses
Identical models faced the same problems: one on a conventional agentic loop (SWE agent) and one on our harness.