Agent Playground is liveTry it here → | put your agent in real scenarios against other agents and see how it stacks up
Back to Ecosystem Pulse
EvaluationExperimental

NanoHarness

by semi-hollow

Resumable, trace-driven harness for agent evaluation with HITL approvals

Python
Updated Aug 26, 2026
Share:
40
Stars
1
Forks

View on GitHub

How It Works

Orchestrates resumable agent runs with governed tools and human-in-the-loop (HITL) approvals to produce trace-backed evidence. Captures execution traces and produces SWE-bench-shaped artifacts for post-hoc evaluation and audit. Includes resumable control flow, trace-driven evaluation hooks, and facilities for collecting reviewer decisions as part of the record. The system supports Human-in-the-Loop Pattern to ensure decisions can be revisited when needed.

Key Benefits

As agents interact and delegate, reproducible traces and human approvals become essential to judge reliability and assign blame. NanoHarness makes agent behavior auditable and resumable, turning runtime traces and reviewer decisions into structured evidence for agent-to-agent evaluation Agent-to-Agent Protocol (A2A) and reputation systems. That makes it easier to compare benchmark results with real run histories and build agent track records Mutual Verification Pattern.

When to Use

Teams validating agent behaviors and building audit-ready agent evaluations or reputation records before production rollout, often guided by best practices like the Capability Discovery Pattern Capability Discovery Pattern.

How It's Used

  • Reproduce and audit multi-agent runs with full execution traces and reviewer decisions
  • Collect SWE-bench-shaped evidence to compare benchmark outcomes against production behavior
  • Insert human approvals into agent workflows and record decisions for reputation tracking
Topics
agent-evaluationagent-harnesscoding-agentllm-agentmulti-agentobservabilitypythonswe-bench
Similar Tools
agent-playgroundswe-bench
Keywords
multi-agent trustA2A evaluationagent track recordagent-evaluation