agent-jit

In an agent loop, multi-step tool orchestration, state maintenance, and intermediate data flow are fundamentally deterministic, yet they still burn tokens and model round trips. Agent JIT lifts those deterministic paths out of the LLM loop, compiles them into a restricted DSL program, and hands the result to a schema-validated graph runtime for deterministic execution. The agent stops paying tokens for steps that do not require intelligence.

The plugin registers two core meta-tools: jit_describe_tools renders any host tool as a deterministic functional contract, and jit_execute_program compiles DSL source into schema-validated IR before graph-runtime execution. Every tool already registered in ctx.tools-bash, glob, web_search, or tools provided by other plugins-is immediately orchestratable via DSL with zero configuration and zero code changes, and execution still goes through the full policy pipeline. Benchmarks show roughly a 9.5x token reduction, 4x fewer model round trips, and 9.1x less intermediate data exposed to the model compared to iterative tool calls, while task correctness remains at 100%.

Built for developers running multi-step workflows in DeepSeek Harness who want to dramatically cut token spend and latency on programmable orchestration paths. The agent remains the planner; Agent JIT is purely an execution offload layer that does not alter the agent's planning role.