Runtime
How Facio moves from request to reviewed action, memory, audit, and resume.
Facio is built around a long-running agent loop. Each turn gathers context, chooses a model/provider route, uses tools where allowed, asks for human review when needed, records what happened, and keeps enough state to continue later.
The important idea is the shape of the runtime: Facio treats reliability as behavior across a whole task, not as a single model response.
The runtime loop
| Step | What happens | Why it matters |
|---|---|---|
| Observe | Facio assembles session state, channel context, memory, files, recent history, and tool state. | The agent does not start from zero on every message. |
| Decide | The runtime routes work through the selected provider, model, skills, tools, and policy boundaries. | Teams can change capability and cost without changing the product flow. |
| Act | The agent can use browser work, files, scripts, MCP, channels, media tools, and other configured capabilities. | Useful work can continue beyond a chat reply. |
| Ask | If a step crosses a risk or authority boundary, Facio pauses for human input or approval. | Review becomes part of the runtime, not a vague prompt convention. |
| Record | Tool use, approvals, errors, usage, and outcomes are kept visible through logs and audit records. | Operators can inspect what happened later. |
| Resume | The task can continue after waiting, approval, retry, restart, or a delayed follow-up. | Long-running work remains useful after interruptions. |
Human review
Facio is designed for workflows where the agent should do work, but not silently take authority. Review can appear as approval, selection, text input, structured forms, or credential capture.
Good review moments answer four operator questions:
- What does the agent want to do?
- Which evidence or context did it use?
- What will change after approval?
- How can the work be paused, rejected, or resumed?
Placet is the default surface for this because it supports structured approvals, file-heavy conversations, streaming progress, and management workflows better than plain chat.
Sessions and memory
Facio keeps session state for the current work and memory for facts that should survive longer. The runtime can remember preferences, durable facts, known issues, history, and lessons from previous work.
That does not mean every detail should become permanent memory. Operators should treat memory as part of the operating model: useful facts go in, sensitive or temporary facts stay scoped to the task.
Tools and boundaries
Tools are powerful because they let the agent inspect, edit, browse, execute, or communicate. They are also the point where authority matters most.
Facio keeps tool use inside configured boundaries:
- File and shell work can run inside a sandboxed environment.
- Credentials are referenced through placeholders instead of being handed to the model as plaintext.
- Channels use explicit allow lists.
- Risky tool chains can be stopped by guardrails or human review.
- Logs and audit records make tool use inspectable.
What to read next
- Integrations explains the provider, channel, MCP, skill, and credential surface.
- Security explains the public defense model.
- Operations explains how to run and maintain a Facio deployment.