You're No Longer Just a Spectator: Real-Time Agent Control in SaaS Factory
You're No Longer Just a Spectator: Real-Time Agent Control in SaaS Factory
v1.0.157 — Collaborative Real-Time Agent Control
Until today, launching an agent cycle in SaaS Factory meant one thing: sit back and watch. The pipeline was autonomous by design, and that was mostly a feature. But "mostly" is doing a lot of work in that sentence.
The fire-and-forget model has a quiet failure mode. An agent interprets a vague spec too broadly. It decides to refactor a module you didn't ask it to touch. It produces a perfectly logical plan that heads in exactly the wrong direction. And because you can't intervene, you wait — then spend time undoing what you could have prevented with a single sentence twenty minutes ago.
Tools like Cursor's Composer, Devin, and Replit Agent Chat have solved this by keeping the human genuinely in the loop. Not as an approver of final output, but as a collaborator who can redirect mid-session. The AI does the heavy lifting; the human provides steering when it matters.
With v1.0.157, SaaS Factory works the same way.
Three controls. One principle.
The new Collaborative Real-Time Agent Control feature gives you three ways to interact with a live pipeline run:
Pause stops the cycle at the next safe boundary, preserving everything the agent has done so far. Resume it whenever you're ready. Nothing is lost.
Steer lets you inject a plain-text instruction directly into the running cycle. The agent picks it up at the next step and uses it as context going forward. Completed steps are untouched. It's the equivalent of leaning over and saying "actually, focus on the payments module" — without blowing up the session.
Override catches agent decisions at checkpoints, before they become pull requests. You can accept, replace, or discard the proposal. Every override is logged, so there's a clear record of where human judgment shaped the outcome.
The underlying principle across all three: autonomous by default, interruptible on demand.
Nothing broke to build this.
One of the cleaner aspects of this release is what it didn't require. The existing Inngest pipeline topology is unchanged. Agent orchestration logic is unchanged. What changed is that the pipeline now checks for interrupt signals at step boundaries, and the UI now surfaces controls to send them.
The pipelineRuns status system — already tracking run state in the database — gains a handful of new flags: paused, resumed, overridden. That's the entire persistence layer. The rest is hooks and UI.
This matters because it means the feature composes cleanly with everything else. Existing pipeline definitions work. Existing observability works. You're not opting into a different mode of operation — you're getting additional controls over the same one.
Full autonomy is still the default.
If you do nothing, runs behave exactly as they did before v1.0.157. The pause, steer, and override controls are there when you need them; they don't impose a mandatory review gate on every step. The pipeline doesn't slow down waiting for human input that isn't coming.
The goal isn't to turn SaaS Factory into a manual process. It's to make sure that when human judgment would genuinely improve an outcome, there's a clean way to apply it — without cancelling the run, losing context, or waiting until after the damage is done.
Collaborative Real-Time Agent Control is available now in v1.0.157. See the feature documentation for the full technical details.