Skip to main content

Planning

Planning mode lets a parent agent plan first, then execute using a visible scratchpad and a strict step loop. It’s ideal for complex, multi-step tasks that need transparency and reliable progress.

⚠️ Incompatibility: When Planning mode is enabled, Voice cannot be enabled for this agent.

When to use​

Turn on Planning when you need the agent to:

  • Decompose a broad request into clear sub-goals
  • Execute step-by-step with tool-call discipline
  • Show live reasoning and progress for debugging/audit
  • Auto-verify each step before moving to the next

Enable in the UI​

  • Open your agent → click Edit.
  • In the Sub-Agents card, click the gear icon next to Add.
  • In the side drawer, set Orchestration type → Planning (options: None, Planning).
  • Save. If Voice was enabled, you’ll be prompted to disable it.
  • Publish.

What it does (runtime)​

1. Planning​

A visible plan created before execution:

  • Overall Objective — one-sentence summary of the user’s goal
  • Sub-Goals List — numbered, actionable steps (1…N)

Benefits: continuity across turns, transparent reasoning, and fewer hallucinations via explicit decomposition. Users can view the plan during and after the run (read-only).

2. Step-Wise Controller (Define → Execute → Evaluate)​

PhaseAgent action
Evaluate & DefineEvaluate if the previous step was successfully executed and then write a precise, current-step target (e.g., “Retrieve customer purchase history from DB”).
ExecuteChoose one action: call a tool or produce a final answer for that step.

What users see (presentation)​

  • Exposed Thought Stream — live updates for each Evaluate & Define → Execute cycle.
    Collapsible sections so power users can expand raw reasoning; sections auto-collapse after the final answer.
  • Tool State Visualization — for each tool call, show:
    • Inputs (e.g., query/params)
    • Status: pending → running → success/failure
    • Output preview (first lines or summary)
    • If using Browser use, a playback panel (live stream where available; otherwise snapshots/GIFs of interactions).
  • Scratchpad Dashboard — renders sub-goals as a checklist; completed steps are ticked ✔ and annotated with short notes.

Best practices​

  • Keep tools scoped & documented so the planner picks the right one.
  • Start with fewer, larger steps; too many micro-steps increase latency.
  • Use global/per-tool limits (Tools → gear) to prevent runaway retries.
  • Pair with Guardrails (policy/PII) and Reflection for format/style checks.
  • Ask the agent to summarize deltas after each step (what changed, what’s next).

Troubleshooting​

  • Plan looks noisy/overlong → tighten the Base instruction (max N steps, keep “Goal” lines short).
  • Wrong tools chosen → improve tool names/descriptions; mention critical tools in Base instruction.
  • Stalls on Evaluate → raise per-tool timeout or relax the step goal; verify tool parameters.
  • Too slow → reduce step count, lower retries, or increase thresholds/timeouts only where needed.

With Planning orchestration, your agent becomes a transparent, dependable executor for multi-step work—planning in the open, verifying each move, and leaving a clear audit trail from intent to outcome.