Steering Running Agents

A live run drifts toward the wrong file. You don't have to let it finish or throw it away — you can redirect it mid-flight and keep every bit of context it has already built.

Why this, for you: Lesson 7 named steer / restart / let-it-finish in one paragraph. This lesson is the operator's manual: which lever to pull, when, and the interface gotcha that makes a "steer" silently turn into a "wait." It's the human-in-the-loop control that the whole harness leaves room for — the place your judgment still moves the run.

A steering message is a mid-run message that interrupts tool execution and redirects behavior without discarding accumulated context. A follow-up is a correction queued during execution and delivered after the current step. Both preserve context; restarting throws it away.

1 Four moves, one decision

When a run is live, you have four options — and reading the tool-call stream is what tells you which.

Move	Use when
Steer	The direction is clearly wrong but recoverable — wrong file, wrong approach, misread scope
Follow up	The current step is fine; you want to adjust the next one ("also update the tests")
Restart	It's too far down the wrong path — a fresh context with a better prompt is cheaper
Let it finish	The approach is acceptable, even if not ideal

Steer as early as possible — ideally right after the first tool call that signals a problem. Indicators: reading files unrelated to the task, creating new abstractions where existing ones should be reused, or repeating the same search with minor variations (stuck).

# you observe in the tool-call stream: read_file("src/payments/stripe_client.py") # wrong module read_file("src/payments/webhook_handler.py") # still wrong # steer NOW — before it reads ten more unrelated files "Stop. Refactor src/auth/ only — session.py and tokens.py. Leave payments alone."

2 Why steering beats restarting

LLM inference is stateless between calls, but context is not — each tool call appends its inputs and outputs to the conversation history resent on the next call. A steering message prepends a correction to everything the agent has already learned: file contents, errors, prior results. A restart discards all of it and forces the model to re-read files and re-derive conclusions, paying token and latency costs again.

That's the whole case for steering: it's the cheapest way to fix a recoverable wrong turn, because the expensive part — the context the agent built — is exactly what it keeps.

3 The interface gotcha

Steering only works if your message actually interrupts. In Claude Code, typed messages queue until the next turn boundary — pressing Enter alone does not stop the current step. To interrupt immediately, press Ctrl+C first, then send. Sending a follow-up when you meant to interrupt lets the wrong step keep consuming context, and the correction arrives too late to be cost-effective. The anti-patterns rhyme: waiting too long (context burned on useless work you still have to undo), and over-steering (interrupting every few steps usually means the initial prompt was underspecified — restart with a clearer one).

When a steer can't save it

Irreversible side effects already executed: a steer redirects future steps but cannot undo a sent request or a pushed commit — which is exactly why Lesson 14's rollback-first design has to come first. Heavily cached sub-agent setups: a steer to the orchestrator doesn't reach already-dispatched workers; only the next dispatch gets the correction. Contaminated context: if earlier tool calls filled the window with verbose errors or irrelevant files, preserving that context hurts more than it helps — restart with a cleaner prompt. Steering is the right tool for a recoverable wrong turn, not for a run that's already poisoned.

↪ Your win: redirect live, keep the context

Read the tool-call stream — wrong-file and stuck-search signals are visible before the output is.
Steer for recoverable drift, follow up for the next step — both preserve accumulated context.
Restart when it's fundamentally wrong — a clean prompt beats salvaging through repeated steers.
Interrupt for real — Ctrl+C first; a typed message alone only queues.
Don't steer past irreversibility — a steer can't undo a sent email or a pushed commit.

Retrieval practice — recall, don't peek

Question 1A steering message differs from a restart in that it…

Question 2You should steer when the agent's direction is…

Question 3In Claude Code, a typed message during a step by default…

Question 4Steering every few steps usually signals that…

Question 5 · spaced recall from Lesson 14Rollback-first design tells you to choose…

Ask me anything. Want a quick rubric for reading your agent's tool-call stream for drift, or to map steer-vs-restart onto a run that keeps going sideways? Next in Part 8: Sandboxing & Blast-Radius Containment — the runtime boundary that holds when a permission rule fails.