The Flow

How every LLM call flows through Splinter

The Core Flow

Every LLM call in Splinter follows this exact path:

Step by Step

1. Agent Calls Gateway

# Agent.run() internally calls:
response = await gateway.call(
    agent_id="researcher",
    provider=LLMProvider.OPENAI,
    model="gpt-4o-mini",
    messages=[...],
    state=shared_state,
)

2. Control Checks

Gateway asks the Control Layer:

# Internal to Gateway
self._control.check_before_call(agent_id, tools)

This checks:

• Budget: Have we exceeded max_budget?
• Steps: Have we exceeded max_steps?
• Time: Have we exceeded max_time_seconds?
• Tools: Is this agent allowed to use these tools?
• Loops: Is the agent stuck in a loop?

If any check fails, an exception is raised and the call is blocked.

3. Route to LLM

If Control says OK, Gateway routes to the provider:

# Internal to Gateway
provider = self.get_provider("openai")
response = await provider.complete(request)

4. Record Metrics

After the call, Gateway records:

# Internal to Gateway
self._control.record_call(response, agent_id, state)

This tracks:

• Cost (tokens × price)
• Token usage
• Latency
• Call history

5. Coordination

Gateway updates coordination:

# Internal to Gateway
self._coordination.checkpoint(
    workflow_id=workflow_id,
    step=step,
    agent_id=agent_id,
    state=state,
)

6. Return Response

The response flows back to the agent, which parses it and continues.

Why This Matters

Safety

Every call is checked. No agent can exceed limits.

Observability

Every call is tracked. You know exactly what happened.

Resumability

Every step is checkpointed. Failed runs can resume.

Coordination

State is shared. Agents work together, not against each other.