Agent harnesses leave the controller code: natural-language policies run on a shared runtime and stay competitive on three benchmarks
Synopsis
An agent's external execution harness is usually entangled with controller code, which makes it hard to inspect, compare and ablate. Researchers write run-level harness policy as editable natural-language documents (NLAHs) that a shared runtime, IHR, interprets into child-agent calls, handoffs, state updates and validation gates; on SWE Verified, Terminal-Bench 2.0 and OSWorld, IHR-executed NLAHs land in the same task-performance regime as code and prompted realizations while the static policy text shrinks sharply.
Interpretation
An agent's external execution harness can be written as an editable natural-language policy document and executed by a shared runtime, instead of being bound to controller code. Harness logic has usually been buried in tightly coupled controller code, mixing prompts, tool adapters, parser rules, validation scripts, retry logic and benchmark-specific assumptions, which makes harnesses hard to inspect, port, compare and ablate. On Live-SWE for SWE Verified, the IHR-executed NLAH reaches 73.0 against 67.0 for the native code harness and 77.0 for the same text used as ordinary prompting; on SeeAct for OSWorld the NLAH reaches 46.3 against 47.1 for code; on MHTBA for Terminal-Bench 2.0 the NLAH reaches 53.9 against 57.3 for prompting and 36.0 for code.
Natural language carries only harness policy, while tool execution, parsing, sandboxing and logging stay in code where precision matters. A fixed runtime-policy instruction turns a minimal base agent with only a terminal into IHR, which launches child agents from the document, passes task packets, supervises handoffs and records behavior. All experiments use Codex CLI version 0.123.0, gpt-5.4-mini and xhigh reasoning effort, running in Docker containers on Ubuntu 24.04 servers with per-task caps of 32 vCPUs, 84 GiB memory and 40 GiB storage.
Perspective
The demonstrated scope is three harness families: Live-SWE on SWE Verified, MHTBA on Terminal-Bench 2.0 and SeeAct on OSWorld, all run under one setting of Codex CLI 0.123.0, gpt-5.4-mini and xhigh reasoning effort; readers should treat it as feasibility evidence for a policy representation layer, not a production deployment recipe.
Natural language is imprecise, so semantically important constraints may be under-specified, interpreted differently across models, or weakened by paraphrase, which is why executed behavior must be checked through runs rather than inferred from text. The prototype runtime's orchestration reliability and information handoff recall fall below the direct-context prompted realization, with handoff recall of 0.322 and 0.553 under parent-child execution. In module ablations, multi-candidate search raises SWE agent calls from 1.1 to 5.7 but drops performance from 73.0 to 71.4, and context compression lowers OSWorld performance from 44.4 to 36.1.
