Skip to main content
The model is the engine. The harness is everything around it that turns a capable-but-erratic engine into a reliable system: the control flow, the guardrails, the retries, the sub-agents, and the evals. A strong harness is often the difference between an agent that demos and one that ships.
The context layer is one component of the harness — a critical one, but not the whole thing. This page covers the rest, and how it wraps the context layer.

​
Components that earn their keep

​
Soft gates against early exits

Agents love to declare victory early — returning a partial answer, skipping a verification step, or stopping before the task is actually done. A soft gate is a checkpoint that asks “are the exit conditions genuinely met?” before the agent is allowed to finish, and sends it back if not. Unlike a hard rule, a soft gate is itself a judgment step — it can reason about whether the work is complete for this task, which is what makes it robust to fuzzy, non-repeated workflows (the same ones where agents otherwise stall — see Drill-down limitations).

​
Critic sub-agents for in-session evals

A second agent whose only job is to attack the first agent’s output before it reaches the user: does the reasoning follow from the retrieved context? Is any claim unsupported? Did it use stale or conflicting knowledge? Running evals in-session — not just offline on a benchmark — catches failures at the moment they happen, when they’re still cheap to fix. Independent critics also counter the single-model failure mode where the agent is confidently wrong and has no mechanism to notice.

​
Retrieval as a tool, not a preamble

The harness should give the agent a search tool and let it discover context, rather than pre-computing context and stuffing it into the prompt. This is both a harness decision and a context-layer decision — covered in Context search accuracy.

​
How the pieces compose

The context layer is the agent’s memory and knowledge; the harness is its judgment and self-control. You need both — a perfect harness over a poisoned context layer still ships confident nonsense, and a perfect context layer under a naive harness still exits early and never checks its work.

​
Where to start

If you’re building a domain agent from scratch, the highest-leverage order is usually:
1

Fix the context layer first

Structured, searchable, attributable context removes the largest class of failures. Start at Why Open Index.
2

Add a critic pass

One sub-agent reviewing outputs against retrieved context catches most of what slips through.
3

Add soft gates on the exit

Stop early exits on the fuzzy, multi-step tasks where autonomy matters most.
4

Close the learning loop

Let the agent write back what it learns. See Self-learning loops.