Verify Before Done

Agents favour their first plausible answer and declare completion early, so the harness must force verification before "done": a pre-completion checklist, end-to-end tests run as a user would, or an agreed sprint contract.

Evidence: moderate

Left to themselves, agents stop too early. LangChain observed that models favour their first plausible answer. Anthropic's long-running harness work found agents marking features done without testing them end to end, and later sessions declaring a whole project complete on the strength of partial progress. A dynamic-workflows post on claude.dev names the same tendency "agentic laziness". An agent's own "done" is therefore weak evidence that anything works, and a self-chosen check is often a proxy that only resembles success.

The fix in every source is to make verification a step the harness enforces rather than one the agent may choose. LangChain added a pre-completion checklist middleware that runs a self-verification loop before the agent may finish, as part of the harness-only changes that took its agent from 52.8% to 66.5% on Terminal-Bench 2.0. Anthropic's long-running harness found that explicitly prompting agents to test through browser automation, as a user would, markedly improved results. Its later application-development harness goes further, with a separate evaluator that drives the live application, grades against hard thresholds and agrees a "sprint contract" of testable completion criteria with the generator before any code is written. Anthropic's Claude Code guidance opens with giving the agent a way to verify its work. Meta's Muse Code builds the check into the harness: a pinned goal names its acceptance checks, and a completion audit refuses to close the goal until they pass (see pinned goal with harness-audited acceptance checks).

The sources are consistent practitioner and lab reports, with LangChain's benchmark gain covering a bundle of changes rather than verification alone.

Verification can be computational, through back-pressure and a runtime the agent can observe, or done by a separate evaluator. For agents that change live systems, the strongest version is that the system or its owner decides, not the agent: a repair is finished when an independent check confirms the system is serving again, as argued in owner validation decides success.

Sources

  • LangChain (Viv Trivedy), 17 Feb 2026 — https://www.langchain.com/blog/improving-deep-agents-with-harness-engineering
  • Anthropic Engineering (Justin Young), 26 Nov 2025 — https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents
  • Anthropic Engineering (Prithvi Rajasekaran), 24 Mar 2026 — https://www.anthropic.com/engineering/harness-design-long-running-apps
  • Anthropic, undated, section 'Give Claude a way to verify its work' — https://www.anthropic.com/engineering/claude-code-best-practices
  • claude.dev, 2 Jun 2026, 'agentic laziness' — https://claude.dev/blog/a-harness-for-every-task-dynamic-workflows-in-claude-code/
  • Meta Model API cookbook, Building with Muse Code, undated (retrieved 3 Oct 2026) — https://dev.meta.ai/docs/cookbook/goal-tracking