Trace Based Agent Harness Generation

via “multi-model agent harness abstraction with swappable implementations”

An Open Agent Computer for ANY digital work.

Unique: Treats Agent Harness as a swappable, pluggable component that abstracts specific LLM implementations and reasoning patterns. Different harnesses can be selected per workspace, enabling multi-model support and experimentation without runtime changes.

vs others: Provides explicit harness abstraction enabling multi-model and multi-architecture support, whereas most agent frameworks are tightly coupled to specific LLM APIs or reasoning patterns.

via “trace-based agent harness generation”

We built meta-agent: an open-source library that automatically and continuously improves agent harnesses from production traces.Point it at an existing agent, a stream of unlabeled production traces, and a small labeled holdout set.An LLM judge scores unlabeled production traces as they stream.A pro

Unique: Generates agent harnesses directly from execution traces rather than from manual specifications, using trace analysis to infer effective prompts, tool selections, and control flow automatically

vs others: Unlike prompt engineering tools that require manual iteration, this learns from successful execution patterns, reducing the feedback loop from hours of manual testing to minutes of trace analysis

via “agent-behavior-testing-harness”

[Interview: About deployment, evaluation, and testing of agents with Sully Omar, the CEO of Cognosys AI](https://e2b.dev/blog/about-deployment-evaluation-and-testing-of-agents-with-sully-omar-the-ceo-of-cognosys-ai)

Unique: unknown — insufficient data on specific tracing implementation (instrumentation approach, trace storage, visualization UI)

vs others: unknown — insufficient data on how testing harness compares to general LLM debugging tools