Mid-Loop Chat Steering: Talk to the AI While It Works

June 28, 2026 · InfraScout Team

A single InfraScout chat turn is rarely quick. When you ask the AI to review a tenant or sweep a fleet of agents, it spends minutes working — calling tool after tool, reasoning over what comes back, and building toward an answer. Until now, once that work started, you were a spectator: you could wait it out or hit Stop, and nothing in between. If you realized halfway through that you wanted to redirect the investigation, your only move was to interrupt and start over, throwing away everything the AI had already gathered.

This release changes that. You can now keep typing and send a message while the model is still working, and InfraScout folds it into the running assessment at the next safe moment — no restart, no lost progress. We also fixed the other half of the problem: pressing Stop no longer kills the conversation. Together these turn an InfraScout chat into something that behaves the way a good CLI agent does — you stay in the driver's seat the whole time the AI is running.

Steer without stopping

The composer is no longer disabled while a turn is streaming. As the AI works through tool calls, you can type your next message and send it right away. InfraScout holds that message until the running turn reaches a safe boundary — the gap between one step and the next — and then injects it into the same conversation the AI is already working in. The AI reads it, takes it into account, and adjusts course, all without abandoning the work it has done so far.

That difference matters most in the middle of a long investigation. Picture the AI a few minutes into a domain-controller review when you spot that it's only looking at one site. Instead of stopping and re-explaining everything, you just type "also check the other DCs in the EU region" and send. The AI picks it up at the next step and widens its scope, keeping every finding it had already collected. The same move works for tightening things down — "focus only on accounts with privileged roles" — or for adding a constraint partway through, like "don't run anything that changes state, read-only from here."

What you see when you send mid-stream

Because your message lands while the AI is busy, it doesn't appear as a normal message right away. You'll see a short-lived pending chip the moment you send — confirmation that InfraScout has your input and is holding it for the right moment. As soon as the AI reaches a safe boundary and takes the message in, that chip turns into a real message bubble in the running conversation, exactly as if you'd sent it at the start of a turn. There's no separate window, no second thread, and nothing to reconcile afterward: your steer becomes part of the same transcript.

TIP

Send one clear instruction rather than a rapid burst of fragments. The AI applies your steer at the next boundary, so a single, well-formed message — "narrow this to the finance OU and skip service accounts" — redirects the run more reliably than three half-thoughts sent in quick succession.

Stop now pauses — it doesn't end the chat

The second improvement fixes a frustration that has been with InfraScout since the beginning. Pressing Stop used to flip the chat into a terminal state: the run halted, but so did the conversation, and the only way forward was to start a brand-new chat. Anything the AI had done up to that point came with you only as a record you couldn't build on.

Now Stop behaves the way Esc does in a command-line agent. When you stop a running turn — or if your connection drops mid-stream — InfraScout preserves the partial turn and leaves the chat active and ready to take your next message. You can read what the AI managed to gather, then pick up right where you left off: refine the question, send a correction, or point it somewhere new. The conversation stays alive, so an interruption is a pause, not a dead end.

INFO

A stopped turn keeps everything the AI produced before you stopped it. Use Stop freely as a steering tool in its own right — halt a run that's heading the wrong way, look at what it found, and send a sharper instruction into the same chat.

Why this changes how you run assessments

Taken together, mid-loop steering and a non-terminal Stop make a long assessment feel like a conversation again rather than a one-shot command you either watch or abandon. You no longer have to write the perfect opening prompt and hope, because you can correct course as the AI reveals what it's finding. When a review turns up something unexpected three minutes in, you react to it immediately instead of waiting for the whole turn to finish and then starting another. When the AI is clearly on the wrong track, you stop it, look, and redirect — without losing the parts it got right.

The result is fewer wasted runs and tighter investigations. Scope-narrowing that used to mean re-running an assessment now happens inline. A mid-course constraint — read-only, this OU only, skip these hosts — lands the moment you think of it. And because every steer and every stopped turn stays in the same transcript, the conversation remains a faithful record of how the assessment actually unfolded.

Try it

Open a chat and ask for something substantial — a tenant security review or a fleet-wide check that you know will run for a while. While the AI is working, type a refinement and send it: watch the pending chip appear and then resolve into a message as the AI takes it on. Try stopping a run partway through, reading what it found, and sending a follow-up into the same conversation. You'll find you can shape an assessment as it happens instead of planning every detail up front.

See the chat guide for more on running assessment conversations. Questions or feedback? Reach us at info@infrascout.cloud.