跪拜 Guibai
← All articles
Agent

Agent Output Quality Depends on the Harness, Not Just the Model

By 飞哥数智谈 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Teams chasing better agent results by swapping models are optimizing the wrong variable. The same model produces order-of-magnitude differences in output quality depending on how tasks are defined, context is managed, and human checkpoints are placed — and those practices are cheap to adopt compared to model upgrades.

Summary

Two people using the same agent can get wildly different results. One gets hollow, formatted text that needs endless correction; the other gets output that increasingly matches real work needs. The difference is the Harness — the working environment built around the model. Task clarity, context curation, tool permissions, human intervention points, result verification, and failure handling collectively determine whether a model's capabilities translate into stable, useful output.

A vague instruction like "analyze the market" is an idea, not a task. An agent will guess, and guess wrong with confidence. Specifying background, objective, expected result, boundaries, and acceptance criteria — and requiring every judgment to cite source evidence — prevents fabrication. Context is not about dumping documents; it's about showing the right information at the right time, including persisting intermediate state so long-running tasks can recover from drift.

Human-agent division shouldn't happen at the job-role level but at the task-unit level. Clear, low-risk, reversible tasks can run autonomously; fuzzy, high-stakes, or irreversible operations need human control. And a single successful run isn't enough — the prompt, context, rules, workflow, and known failure modes should be packaged into a reusable capability that gets retested on a different instance to avoid overfitting to one case.

Takeaways
Agent performance gaps between users stem from the Harness — task definition, context, tools, human checkpoints, and failure handling — not model instability.
A vague request like "organize customer info" is an idea, not a task; an agent will guess and confidently produce wrong output unless given background, objective, expected result, boundaries, and acceptance criteria.
Requiring every judgment to cite source evidence and marking unconfirmable content as "pending human judgment" prevents agents from fabricating details.
Context is not about volume — dumping dozens of documents causes the agent to miss what matters; the goal is showing the right information at the right time.
For long-running tasks, persisting intermediate state (completed steps, conclusions, remaining issues) lets the agent recover from drift without restarting from scratch.
Human-agent division should happen at the task-unit level, not the job-role level: clear, low-risk, reversible tasks suit autonomy; fuzzy, high-risk, irreversible operations need human control.
Deleting databases, making payments, and publishing externally are operations where saved time is not worth the risk of automation errors.
A successful agent run should be solidified into a Task Capability Pack containing the prompt, task definition, context, execution flow, evaluation rules, and known failure modes.
Retesting a solidified method on a different but similar task catches overfitting — a method that only works on the original case hasn't truly generalized.
Conclusions

Framing agent reliability as a Harness problem rather than a model problem shifts the optimization target from expensive model swaps to cheap process design, which most teams underinvest in.

The advice to persist intermediate state during long agent runs mirrors checkpointing in distributed systems — a pattern that agent workflows have not yet standardized on but clearly need.

Requiring source citations for every agent judgment is a lightweight guardrail that costs almost nothing to implement yet directly attacks the hallucination problem at the output layer rather than the model layer.

The overfitting analogy for agent workflows is sharp: a prompt that works once may have succeeded due to accidental context, and without retesting on a different instance, teams are just accumulating brittle templates.

Concepts & terms
Harness
The scaffolding around a large model that makes it capable of completing real tasks: task definition, context management, tool and permission configuration, human intervention points, result verification, and failure recovery. The formulation is Agent = Large Model + Harness.
Task Capability Pack
A reusable bundle that captures everything needed to reproduce a successful agent run — not just the prompt, but the task definition, required context, execution workflow, evaluation rules, and known anomalies with their resolutions.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗