Agent OS, Part 2: The harness
Stages 3 and 4: the control surface around the model, a stop on the loop, then trace and eval before anything ships.
How it works
Day ~185 · Jun 2026 Put a harness on it, and a stop on the loop
Harness is a buzzword until you draw it. It's the control surface around the model (tools, connectors, permissions, defaults) so the horse doesn't run random.
The loop is the other half: call tools until an end-loop guardrail fires, then reply. Without a stop condition you get permission stalls and actions that keep firing. When the ending is unclear, ask the user: do all of them, or only list them. If the agent is stuck on a permission, it should notify you instead of sitting silent.
Day ~190 · Jul 2026 Trace it, eval it, then ship
A harness without health checks is incomplete. Any agent that claims to improve itself without an eval has a gap.
One agent run is the user's ask, the tool loop, then the reply. Trace that tree: retrievals, tool calls, latency, tokens. Evaluate health and quality. Diagnose. Then ship a better prompt or config, or fix the bug and re-run.
My take
What I learnedA harness rides the LLM horse, and every loop needs an end condition.
What I learned"Self-improving" without eval is cosplay.
Try this
- List the tools your agent may call. Anything irreversible (send, pay, publish, delete) sits behind an approval. Prep is automatic, the action waits. Then write the end-loop rule in one sentence.
- Keep one dashboard or log that answers two questions. Was this run healthy, and was it good? Hold ship until Pass.
Sources
- Building effective agents (Anthropic, Dec 2024). Tools in a loop, stopping conditions, human checkpoints.
- Tracing (OpenAI Agents SDK docs). Traces and spans for model calls, tool calls and guardrails.
- Define success criteria and build evaluations (Anthropic docs). Measurable criteria and automated grading.