Skip to main content
Agents in reasoning mode process every request through a pipeline of flow steps. The entire pipeline is exposed on the agent’s Configuration tab — there are no hidden system instructions. Each LLM-powered step is a card you can open, inspect, and customize: what you see in the pipeline is exactly what runs.

​
The pipeline

StepRequiredWhat it does
OrchestrateYesPlans the next move — reviews the request, current state, and prior results, then picks one capability to run.
ActionYesWhere your capabilities execute (API calls, LLM calls, workflows, webhooks, sub-agents). Not an LLM step — manage these on the Capabilities tab.
ProcessOptional — on by defaultDistills each action result into a rolling summary and findings, and writes state updates. Two sub-calls: Summary & Findings and Form Updates. Disable it to keep a raw message transcript instead.
Progress ProbeOptional — off by defaultAfter each result, checks whether the agent should continue, respond, or rewind. Adds a guardrail LLM call per iteration.
RespondYesProduces the final, user-facing answer once the loop decides the task is done. Freeform text by default; you can bake in a structured JSON schema on the step (Response Format in the step config), and callers can still override per request via response_format — see Agent Call — structured final response.

​
Configuring a step

Click any step in the pipeline to open its configuration:
  • System prompt — every step ships with Maitai’s default prompt template. Edit it to tailor the step to your agent; a badge on the step shows Default vs Customized, and you can reset back to the default at any time.
  • Model & parameters — each step runs as its own intent, so each can use its own model and hyperparameters. A common pattern: a fast, cheap model for Orchestrate and Process, and a premium model for Respond.
Guidance that previously lived in “special instructions” now belongs directly in the step prompts:
  • Prioritization logic, checklists before calling tools → Orchestrate
  • How to interpret and synthesize capability results → Process
  • Tone, formatting, and response policy → Respond
  • Structured JSON answer shape (schema) → Respond → Response Format
Flow steps run through your Application like any other intent, which means each step shows up in request monitoring and can be independently fine-tuned over time.

​
Toggling optional steps

Process and Progress Probe have switches directly on the pipeline. Turning Process off trades context quality for latency; turning Progress Probe on trades latency for tighter loop control.

​
Per-request overrides (SDK)

You can toggle flow steps for a single request without touching the published agent — for example, skipping the Progress Probe on a latency-sensitive call: 
agent = client.agent.get("YOUR_AGENT_NAME", version="prod")
agent.config.system_steps.progress = False

request = agent.to_request(messages=[...], session_id="YOUR_SESSION_ID")
response = client.agent.completions.create(request=request)
The step keys are orchestration, processing, progress, and respond. Overrides are validated against the published version and only whitelisted toggles apply — see Agent Call — per-request overlays.

​
Next