Collaborative Real-Time Agent Control: Pause, Steer, and Override Mid-Run
Collaborative Real-Time Agent Control: Pause, Steer, and Override Mid-Run
Available from: v1.0.157
SaaS Factory's autonomous pipeline now supports real-time human collaboration during active agent runs. You can pause a cycle, inject steering instructions mid-flight, or override an agent decision before it becomes a pull request — without disrupting accumulated context or pipeline integrity.
Why This Matters
Autonomous pipelines are powerful, but a fire-and-forget model has a real cost: if an agent heads in the wrong direction, you either wait for it to finish and clean up afterwards, or you cancel the whole run and lose progress. Real collaborative AI tools — Cursor's Composer, Devin's shell/browser takeover, Replit Agent Chat — have demonstrated that the right model is human-in-the-loop on demand: fully autonomous by default, interruptible when needed.
Version 1.0.157 brings this model to SaaS Factory.
Controls
Pause
Stop an active agent cycle at the next safe Inngest step boundary.
- The run enters a
pausedstate immediately after the current step completes. - All accumulated context, completed steps, and intermediate outputs are preserved.
- The run can be resumed at any time — it picks up exactly where it left off.
When to use: You notice the agent is about to explore a direction you want to reconsider, or you need to supply additional information before it proceeds.
Steer
Inject a plain-text instruction into a live (or paused) run.
- Your instruction is surfaced to the active agent at the next Inngest step hook as additional context.
- Previously completed steps are unaffected — only upcoming steps receive the new context.
- Multiple steering instructions can be added sequentially throughout a run.
When to use: You want to narrow the agent's scope, add a constraint, or redirect its focus without cancelling the run.
Example steering instruction:
Focus only on the payments module. Do not modify the authentication layer.
Override
Replace or discard an agent's proposed decision before it is committed as a PR.
- When an agent reaches a decision checkpoint, a confirmation prompt appears in the pipeline UI.
- You can accept the proposal, replace it with your own directive, or discard it entirely.
- All overrides are logged against the pipeline run for full traceability.
When to use: The agent has produced a plan or decomposition you disagree with, and you want to substitute your own before code is written.
How It Works
The implementation builds on existing SaaS Factory infrastructure:
| Layer | Mechanism |
|---|---|
| State | pipelineRuns table gains status flags: paused, resumed, overridden |
| Orchestration | Inngest step-level hooks check for interrupt signals before each step executes |
| UI | Active run cards expose Pause, Steer, and Override controls |
No changes are made to the underlying pipeline topology or agent orchestration logic. The autonomous default behaviour is fully preserved — these controls are opt-in during a run.
Auditability
Every intervention is recorded:
- Pause / Resume events are timestamped on the
pipelineRunrecord. - Steering instructions are stored as ordered log entries associated with the run.
- Overrides capture the original agent proposal alongside the human replacement, with a timestamp and actor reference.
This makes it straightforward to review what happened in any given cycle and understand where human input shaped the outcome.
Behaviour Summary
| Scenario | Behaviour |
|---|---|
| No intervention | Fully autonomous — identical to pre-v1.0.157 |
| Pause then resume | Run continues from exact pause point; no context lost |
| Steer during run | Upcoming steps receive injected context; completed steps unchanged |
| Override at checkpoint | Agent proposal replaced or discarded; override logged |
| Cancel (unchanged) | Terminates run; no resume possible |