All News
NewsAgent Control

OpenClaw's New Steer Command Lets Operators Redirect AI Agents Mid-Conversation

Until now, correcting a running OpenClaw agent meant waiting for it to finish or killing the session entirely. A new /steer command changes that: operators can inject real-time guidance into active AI sessions, redirect subagents individually, and target ACP harness processes — all without interrupting execution.

May 4, 20264 min read

In-Flight Course Correction

The /steer command targets the active run in the current session. It injects guidance into what's already executing — it won't launch a new run if the session is idle.

Subagent Targeting

Child processes get their own syntax: /subagents steer <id>. Operators can redirect a specific subagent without affecting the primary session or other running children.

ACP Session Steering

Agent Control Protocol harness sessions use /acp steer --session to target those specific environments. External orchestration tools can now course-correct without rebuilding sessions.

Steer vs. Queue

/steer is for immediate course corrections. /queue steer changes default behavior for future messages. The distinction finally has official documentation.

Why This Matters for Production Deployments

Anyone running AI agents in production has experienced the moment: a model starts heading down the wrong path — maybe it's misinterpreting context, pursuing an irrelevant tangent, or about to take an action you didn't intend. Previously, OpenClaw operators had two options: let it finish and deal with the consequences, or terminate the session and lose all context.

The /steer command introduces a third option. It's conceptually similar to a co-pilot tapping the pilot on the shoulder — the agent receives the guidance and incorporates it into its active reasoning without breaking stride. The session state, accumulated context, and in-progress tool calls remain intact.

The Subagent Dimension

OpenClaw's architecture supports spawning child agents for subtasks — research threads, parallel tool invocations, long-running computations. Before today, operators had coarse-grained control: steer the top-level session and hope it propagated, or wait for children to finish.

The new /subagents steer <id|#> <message> syntax provides surgical precision. If one subagent is heading the wrong direction while three others are fine, you can correct just that one. In complex multi-agent workflows, this is the difference between a minor course correction and restarting an entire pipeline.

Loop Detection Gets Smarter Defaults

In a separate but related documentation update, OpenClaw has narrowed its loop detection recommendation. The previous guidance suggested enabling it universally. The new position is more nuanced: smaller models should start with loop detection enabled, while flagship models — which rarely enter repetitive loops — can leave it disabled.

This reflects a practical reality that operators have observed: large frontier models almost never get stuck in tool-call loops, and the detection overhead (plus occasional false positives) isn't worth it. Smaller, faster models — the ones operators often use for high-volume, lower-stakes tasks — benefit more from the safety net. It's a small documentation change, but it signals OpenClaw paying attention to how different model tiers behave in production.

The Bigger Picture: Operational Maturity

Both changes point in the same direction: OpenClaw is building out the operational control surface that production AI deployments need. Real-time steering gives human operators a direct feedback channel to running agents. Tiered loop detection acknowledges that one-size-fits-all guardrails create unnecessary friction. Together, they represent a platform that's learning from how operators actually run AI systems at scale — not just how developers build them in development.

Real-time agent control. Already deployed.

DeployClaw ships every upstream OpenClaw update the moment it lands — including the steer command. Your agents are already steerable.

DeployClaw News · by Carlos Simpson

DeployClaw hosts OpenClaw instances. Upstream fixes ship automatically. This publication covers development independently.