Harnesses

A harness executes the agent's instructions, tools, skills, and runtime logic. Its FlashRSI adapter declares which resources belong in a Snapshot, then runs the selected resource version through its execution entry point.

Available integrations

Harness	Editable resources	Runtime and details
MiniAgent	Instructions, skills, tools, configuration, declared Python code	Direct mapping of a small native Python harness. Usage · source
Claude Code	Instructions, skills, commands, agents, hooks, MCP/LSP registrations, settings	Official Agent SDK (claude-agent-sdk>=0.2.153,<0.3); exposed resource interfaces are mapped to files. Adapter details
Codex	Python core, instructions, tools, skills, agents, hooks, plugins, memory, rules/config files	Snapshot-local Python semantic port based on rust-v0.147.0; runs in an isolated process. Guide · version and migration boundary

Codex's Python core implements the documented paths from a pinned Rust baseline. It does not launch the Codex CLI, fall back to the host core, or claim complete equivalence to every upstream feature. For example, declaring rules as resources does not implement rules-based execution policy; see capability limits.

Target and Agent Worker

The same adapter can solve benchmark tasks as a Target or edit candidate files as a Worker. Each role receives its own configuration and model connection:

agent:
  name: miniagent
  llm:
    model_name: gpt-5-mini
algorithm:
  worker:
    type: agent
    name: claude_code
    llm:
      model_name: your-worker-model

This fragment only illustrates role selection. Keep the rest of the destination method's example and supply both roles' credentials and endpoints.

Upcoming adapters

Harness	Status
Pi	Doing
Hermes	Doing
OpenCode	Doing

These are planned integrations and are not available CLI harness choices.

Check the task environment

The harness needs the tools the task requires. Container execution also needs a compatible Linux environment and explicit model credentials. Benchmark data, simulators, Docker images, and graders are supplied by the benchmark integration.

See harness / benchmark responsibilities and Add a harness.