Human-in-the-Loop (HITL) Planner
Map your agent's steps, choose which ones need human approval, and export an interception plan with a textual flowchart.
Interception Plan
| Step | Step Name | Human Approval | Approval Type | Timeout | Fallback |
|---|
Textual Flowchart
Generate the plan to see the step → approval gate → continue/fallback flowchart.
Professional Insights & Guide
Learn how HITL interception points are designed, how gates and fallbacks combine, and where the pattern breaks.
Core Use Case scenario
Automation teams design agent workflows where most steps run autonomously but a few high-stakes actions pause for a human. Each step either passes through without intervention or stops at an approval gate where a person approves, rejects, or edits the proposed action. If nobody responds within the timeout, a fallback behaviour keeps the system safe: the run aborts, retries, skips the step, or continues with a note. The planner makes those decisions explicit before the workflow is built.
How the Plan Is Constructed
- Each step has a name and a requires-approval flag; unchecked steps simply continue to the next one.
- Approval types: Approve/Reject, Edit & Approve, or Assign Reviewer for a second opinion.
- Timeout in minutes defines how long a gate waits before the fallback fires.
- Fallback behaviors: abort the run, continue, skip to the next step, or retry the step.
- The flowchart renders step → approval gate → continue/fallback for every step.
Troubleshooting & Edge-Case Failure Points
- Empty step names: the plan refuses to generate and flags the offending row with an inline error.
- Zero timeouts: a gate with a 0-minute timeout fires immediately, which is rarely what was intended.
- No fallback: without one, a rejected or timed-out gate leaves the workflow undefined and the run can hang.
- Over-gating: flagging every step for approval removes the benefit of automation and stalls throughput.
- Destructive actions: steps like refunds, deletes, or external sends should default to abort on timeout.
Detailed Step-by-Step Instructions
- Name your agent or workflow in the top field, then review the example steps already loaded.
- Edit each step's name, and tick Requires approval only for steps that genuinely need a human.
- For gated steps, choose the approval type, a realistic timeout, and the fallback behaviour.
- Press + Add Step to insert more steps or Remove on a row to delete it.
- Press Generate Interception Plan to rebuild the table and the textual flowchart.
- Press Copy Plan Summary to paste the plan into a design doc, ticket, or PR description.
Related Web Utilities (Silo Hub)
Informative Guides & Helper Articles
How to Use the HITL Planner for AI Agents
Designs human-in-the-loop approval gates: which agent steps pause for review, what fallbacks fire on timeout, what escalates.
- List your agent steps in execution order.
- Mark which need approve/reject/edit gates, timeouts, fallbacks.
- Export the interception plan with flowchart.
Where the Human Belongs in the Loop
Gate placement principle: pause where actions are irreversible, costly, or ambiguous - payments, emails, deletes, production changes. Three defaults that survive reality: every gate needs a safe timeout fallback (deny or queue - never auto-approve); edit-gates beat approve-gates for borderline cases; gate logs become training data for auto-approving the boring 80%. Latency math: a step gating 500x/day belongs in automation with sampling, not a hard stop. HITL is transitional architecture - its goal is its own shrinkage.
HITL Planner for AI Agents FAQ
Which steps need human gates?
Irreversible, costly, or ambiguous ones: payments, external comms, deletions. Everything else runs with logging and sampling.
What happens when a gate times out?
Fail safe: deny or queue, never auto-approve. Timeout auto-approval makes the human a rubber stamp.
How do I reduce gates over time?
Log decisions; where humans approve >99% consistently, auto-approve with audit sampling and outlier alerts.