human-gate-designer

Human Gate Designer

Designs human-in-the-loop review points in DAG workflows: what to present, how to collect feedback, how to route decisions back into the DAG.

---

When to Use

✅ Use for:

• Deciding WHERE in a DAG to place human gates
• Designing WHAT the human sees at each gate
• Defining HOW feedback routes back (approve/reject/modify)
• Balancing automation speed with human oversight

❌ NOT for:

• Runtime execution of human gates (use dag-runtime + Temporal signals)
• General UI/UX design (use design skills)
• Chatbot conversation flow (different pattern)

---

Gate Placement Decision Tree

flowchart TD
  A{Is the action irreversible?} -->|Yes| G1[Gate BEFORE the action]
  A -->|No| B{Is output user-facing?}
  B -->|Yes| G2[Gate AFTER generation, BEFORE delivery]
  B -->|No| C{Cost > $0.50 for remaining nodes?}
  C -->|Yes| G3[Gate at the cost threshold]
  C -->|No| D{Confidence score < 0.7?}
  D -->|Yes| G4[Gate on low-confidence outputs]
  D -->|No| N[No gate needed]

Where to Place Gates

Situation	Gate Position	Why
Irreversible action (deploy, send email, submit)	Before the action	Can't undo
User-facing deliverable (report, website, PR)	After generation, before delivery	Quality check
High cost remaining (>$0.50)	Before expensive phase	Budget confirmation
Low confidence output (<0.7)	After the uncertain node	Expert judgment needed
Ambiguous task decomposition	After planning, before execution	Validate the plan
First run of a new template DAG	After each phase	Build trust gradually

---

Gate Presentation Design

What the Human Sees

┌──────────────────────────────────────────────────────┐
│  🔍 Human Review: [Node Name]                        │
│                                                      │
│  Context: [1-2 sentences: what happened so far]      │
│                                                      │
│  Output to Review:                                   │
│  ┌──────────────────────────────────────────────────┐│
│  │ [The node's output, formatted for readability]   ││
│  │ [Key decisions highlighted]                      ││
│  │ [Confidence: 0.82]                               ││
│  └──────────────────────────────────────────────────┘│
│                                                      │
│  Cost so far: $0.08 / $0.50 budget                  │
│  Remaining nodes: 4 (est. $0.12)                    │
│                                                      │
│  [✅ Approve]  [✏️ Modify]  [❌ Reject]              │
│                                                      │
│  If modifying, what should change?                   │
│  ┌──────────────────────────────────────────────────┐│
│  │ [text input for human feedback]                  ││
│  └──────────────────────────────────────────────────┘│
└──────────────────────────────────────────────────────┘

Presentation Principles

1. Show context, not just output: The human needs to understand what the DAG has done so far, not just the current node's result.
2. Highlight decisions: Bold or annotate the choices the agent made. These are what the human is actually reviewing.
3. Show confidence: If the agent was uncertain, say so. Low-confidence outputs need more scrutiny.
4. Show cost: The human should know what they've spent and what's remaining.
5. Make "Modify" easy: A text input for feedback that gets injected into the retry prompt.

---

Feedback Routing

flowchart TD
  H[Human decision] --> A{Decision?}
  A -->|Approve| C[Continue to next wave]
  A -->|Modify| M[Re-execute node with human feedback injected]
  M --> V[Validate modified output]
  V --> H
  A -->|Reject| R{Reject scope?}
  R -->|This node only| RN[Re-plan this node with different approach]
  RN --> H
  R -->|Entire phase| RP[Re-plan from last successful phase]
  RP --> H
  R -->|Abort DAG| AB[Stop execution, return partial results]

Feedback Injection

When the human selects "Modify," their text becomes part of the re-execution prompt:

Original task: [same as before]
Previous output: [the output the human rejected]
Human feedback: "[the human's modification text]"

Revise your output to address the human's feedback.
Preserve the parts they didn't comment on.

---

Anti-Patterns

Gate After Every Node

Wrong: Requiring human approval after every single node. Right: Gate only at irreversible actions, user-facing outputs, and low-confidence decisions. Most internal nodes need no gate.

Binary Approve/Reject Only

Wrong: The human can only approve or reject, with no way to provide specific feedback. Right: Always include a "Modify" option with a text input for targeted feedback.

No Context in the Gate

Wrong: Showing the human a raw JSON output with no explanation. Right: Show: what the DAG is doing, what happened so far, what this output means, what happens next if approved.