Every time your agent reaches for a real tool, Invoke intercepts the call and runs it through a controlled execution loop before any side effect lands. The loop has three steps and four underlying primitives.
Invoke does not replace your agent framework. It controls what happens when the agent reaches for a real tool. Keep your LangChain chains, Claude Code tools, OpenAI functions, MCP servers, and custom scripts unchanged.
The three-step execution loop
Agent sends intent
Your agent calls Invoke with everything needed to authorize and control the action: the tool name, parameters, a stable agent ID, workflow context, and an optional idempotency key. Invoke controls execution
Invoke runs the call through four sequential controls:
- Validates schema and scope — confirms the tool is within the agent’s registered policy
- Classifies the tool — determines the safety class to decide execution behavior
- Reconciles unknown outcomes — checks live state before allowing a retry
- Gates risky actions — pauses approval-bound tools and blocks failed checks
If any control fails, Invoke stops execution and returns the reason. No side effect lands. Your system gets an outcome
The agent receives a structured result: The four control primitives
Before execution, Invoke assigns a safety class: idempotent, write-once-checkable, irreversible, or approval-bound. The class determines what Invoke does if the first attempt returns an unknown outcome.
Write the commit record
Before the tool touches the outside world, Invoke writes a durable execution record with the execution ID, tool name, parameters, agent ID, and idempotency key.
Reconcile before retry
When an outcome is unknown, Invoke checks live source-of-truth state before allowing another call. If the action already succeeded, Invoke blocks the retry. The agent never needs to implement this logic itself.
Return execution proof
Every completed call returns a structured trace with policy decisions, attempts, reconciliation checks, and the final outcome.
