> ## Documentation Index
> Fetch the complete documentation index at: https://open-index.io/llms.txt
> Use this file to discover all available pages before exploring further.

# What is human steering, and how to reduce it

> Human steering is the hidden tax on agent autonomy. What causes it, why it caps value, and the context and harness changes that reduce it.

**Human steering** is every moment a person has to step in mid-task to redirect an
agent — clarifying an ambiguous instruction, pointing it at the right document,
approving a step it should have taken itself, or nudging it past a point where it
stalled. A little is fine. A lot is the difference between an agent that saves time
and one that just relocates the work.

## Why steering is the real bottleneck

When human steering is required for a task, it caps the **autonomy, speed, and value**
the agent delivers inside the workflow the human was trying to hand off. You wanted to
delegate the task; instead you're babysitting it. Three costs stack up:

* **Throughput** — the agent runs only as fast as a human can attend to it.
* **Scale** — one operator can supervise a handful of agents, not hundreds.
* **Trust** — a system that needs constant correction never earns the autonomy to run
  unattended, so it stays stuck at "supervised."

The value of agents shows up specifically when you cross from *steered* to
*autonomous*. See the [autonomy spectrum](/guide/what-are-autonomous-agents).

## Where steering comes from

Most steering traces back to a handful of root causes — and notably, most are
**context and harness** problems, not model problems:

| The human steps in because…                     | Root cause                                              | Fix                                                       |
| ----------------------------------------------- | ------------------------------------------------------- | --------------------------------------------------------- |
| "That's the wrong document / not what I meant"  | Bad retrieval; guessed context                          | [Context search accuracy](/guide/context-search-accuracy) |
| The agent drifted onto an unrelated instruction | [Context poisoning](/guide/context-poisoning)           | Structured, narrow retrieval                              |
| It stalled on a fuzzy multi-step decision       | Weak harness; no drill-down                             | [The harness](/guide/the-harness)                         |
| "I don't trust this — show me why"              | No [provenance](/guide/provenance)                      | Per-fact attribution                                      |
| It acted on stale/contradictory info            | [Conflicts & staleness](/guide/conflicts-and-staleness) | Per-concept updates + decay                               |

## How to reduce it

<Steps>
  <Step title="Make discovery the agent's job">
    The most common steering — "you're looking at the wrong thing" — disappears when
    the agent searches structured context itself instead of relying on a human (or a
    guessed pre-injection) to hand it the right document.
  </Step>

  <Step title="Contain drift at the data model">
    Structured records and typed edges stop the agent wandering into unrelated
    instructions, which removes a whole class of "why did it do that?" interventions.
  </Step>

  <Step title="Let the harness handle the hard step">
    Soft gates keep it from exiting early on fuzzy tasks; critic sub-agents catch bad
    reasoning before a human has to. The agent recovers on its own instead of waiting
    for a nudge.
  </Step>

  <Step title="Give the human a reason not to intervene">
    Provenance lets the human verify *after* the fact instead of supervising *during*
    it — moving them from the inner loop to the outer loop.
  </Step>
</Steps>

<Note>
  Reducing steering isn't about a smarter model. It's about removing the reasons a human
  *has* to intervene — most of which live in the context layer and the harness around
  the model.
</Note>

<Card title="Next: Determinism" icon="scale-balanced" href="/guide/determinism">
  Making the agent's behavior repeatable enough to trust unattended.
</Card>