How the Agent Works
When you hand Silicon Worker a task, it doesn't just reply with a sentence — it works the task to completion, like an employee would. Understanding this process helps you assign tasks better, follow progress, and step in at the right moments.
The plan → act → check loop
Internally, Silicon Worker advances a task through a loop:
- Plan: based on your task and the current situation, decide what to do next.
- Act: call the right tool to actually do it — read/write files, open web pages, operate apps.
- Check: look at the result, judge whether the goal is met and whether to adjust.
It then takes the new result back to step one and continues, round after round, until the task is done or it needs your input. For more complex tasks, it will first propose a step-by-step plan for you to confirm before starting.
Advancing over multiple turns
A task often can't be done in a single step. Silicon Worker works over multiple turns: do a bit, check a bit, then continue — getting closer to the goal each round. When a single round contains several independent operations, it may also run them in parallel for efficiency.
INFO
You can watch each step in the session view — which tool it called and what result it got. The whole process is transparent.
Long context is compacted automatically
On long-running tasks or lengthy conversations, the accumulated content can exceed what the model can handle at once. When that happens, Silicon Worker automatically compacts the earlier context: it summarizes the past into a more compact form, keeping the key information while freeing up room to keep going. You usually don't need to do anything.
It retries on errors, and falls back when needed
If a model call hits a transient error like a network hiccup, rate limit, or timeout, Silicon Worker retries automatically instead of giving up.
If you've configured a fallback model, then when the primary model keeps failing it will automatically switch to the fallback to keep working, improving overall reliability. See Connect your first model for how to set this up.
High-risk actions pause to ask you
Not every action runs automatically. For high-risk operations that could have real impact (such as writing/overwriting files, running commands, submitting forms, or deleting calendar items), Silicon Worker pauses to ask for your consent first, and continues only after you confirm.
During a task it may also pause and wait for your input because it needs you to:
- Answer a question before it can proceed.
- Confirm a plan it has proposed.
- Approve a high-risk operation.
Once you confirm, it resumes from where it paused — no need to start over. You can also stop a running task at any time. For details on risk levels and permission modes, see Tools & risk levels.
Next / Related
- Skills: teach the agent the standard way to do a class of task.
- Experts & Teams: split complex tasks across collaborating roles.
- Tools & risk levels: the safety mechanism behind pausing and confirming.
- Permissions: configure when it should ask for your consent.