Mid-Stream Steering
Talk to the AI while it works
Don't stop the agent — steer it. While the AI is streaming, the user can drop quick directives ("shorter", "less formal", "go deeper on the second point"). The output keeps flowing, but with the new direction baked in, marked inline so the pivot is visible. No restart, no wasted tokens.
Framing
The problem
Once an agent starts generating, the only intervention is a hard stop — and stopping throws away whatever was already produced.
The pattern
Let users redirect the in-flight stream with quick chips or natural language; the output continues with the new direction marked inline.
Why chat breaks here
Chat treats generation as atomic — you wait, read, and then ask for a redo. Mid-stream steering lets the redirection happen during the work, not after it.
Risks
Too many steering options can fragment output and produce zigzag voice; markers can clutter the result.
Avoid when
Outputs are short enough that waiting and redoing is cheaper than building a steering surface.
Use when
Generations are long enough that waiting and redoing costs more than redirecting in flight.
DOPE evaluation
- Directability
- Direct manipulation of in-flight generation — change tone, depth, or angle without stopping
- Observability
- Course-correction markers appear inline so the pivot is visible in the output itself
- Predictability
- Each steering chip describes the new direction in plain language before applying
- Explainability
- Every steered segment is annotated with the directive that caused it
In the wild
- ChatGPT Voice Mode (OpenAI) — Advanced Voice Mode supports mid-sentence interruption to redirect — the cleanest consumer expression of soft mid-stream steering, well-documented in 2026.
- OpenAI Deep Research mid-run Update (OpenAI) — Feb 2026: users add new info mid-research via sidebar "Update" button. No restart, progress preserved. The non-destructive pivot the pattern actually describes.
- Claude Code mid-turn messages (Anthropic) — Sep 2026 docs: a message typed while Claude runs tool calls reaches it as soon as those calls finish, inside the same turn — the soft pivot, no restart. The steered turn shows the message where it landed; there is no per-segment course marker.
- OpenAI Codex Steer (OpenAI) — Captured Sep 2026: Steer pushed “Skip mercury baths, and name the inventor for each step” into a running five-search task. It lands in the timeline between two searches, the agent acknowledges it and keeps going — no restart. In a pure text stream with no step boundary the same message only lands when the turn ends. The Responses API ships the primitive for builders as response.steer.
FAQ
When should I use the Mid-Stream Steering pattern?
Generations are long enough that waiting and redoing costs more than redirecting in flight.
When should I avoid the Mid-Stream Steering pattern?
Outputs are short enough that waiting and redoing is cheaper than building a steering surface.
What problem does Mid-Stream Steering solve?
Once an agent starts generating, the only intervention is a hard stop — and stopping throws away whatever was already produced.
Why is chat the wrong fit for this?
Chat treats generation as atomic — you wait, read, and then ask for a redo. Mid-stream steering lets the redirection happen during the work, not after it.
Related patterns
- Alternative to: Editable Plan — Steer the plan before it runs vs steer the run while it happens. Different moments, same Directability principle.
- Alternative to: Approval Gates — Hard pause at checkpoints vs soft redirect while flowing. Mid-stream is the gradient inside the binary.
- Often paired with: Reasoning Trace — Watching the reasoning live makes steering meaningful — you redirect what you understand, not what you guess at.