Case study · Architect · 40 min

Agentic Workflow with Human Approval Gates

Orchestrate a multi-step business agent that drafts actions, pauses for human approval on irreversible steps, then resumes reliably.

agentsHITLworkflows

Interview opener

Agent proposes actions that need human approval in enterprise.

1.How do you represent pending actions durably?
2.Approver is offline — timeout policy?
3.Audit trail for regulators?

Strong signals

Explicit state machine for approvals
Human-in-the-loop as first-class, not afterthought
Immutable audit log of proposals and decisions

Hit 0/3

Weak signals (avoid)

Fire tools then ask forgiveness
No durable state for approvals

Requirements

Workflow: intake → research → draft email/CRM update → approve → execute
Human SLA: approvals within 4 business hours; reminders at 2h
Idempotent execution; full audit; resume after deploy
Scale: 2,000 workflows/day; 30% need approval

Back-of-envelope Estimation

Metric	Calculation	Result
Workflows/day	2,000	2,000
LLM calls/workflow	~5	10,000
Tokens/day	10K × 2.5K	25M (~$50)
Approvals/day	0.3 × 2,000	600
Approver load	600 × 2 min	20 hours
Queue storage	state machines × 30d	modest

API Design

Endpoint	Method	Description
POST /v1/workflows	POST	Start workflow with input payload
GET /v1/workflows/{id}	GET	State, history, pending approval
POST /v1/approvals/{id}/decide	POST	{ approve\|reject\|edit, notes }
POST /v1/workflows/{id}/resume	POST	Internal worker resume after approval

Data Model

Entity	Key fields	Notes
workflows	id, type, state, cursor, tenant_id	Durable state machine
steps	workflow_id, name, input, output, status	Append-only log
approvals	id, workflow_id, reviewer, decision	Blocks execute step
outbox	event_id, payload, sent	Reliable side effects

High-level Design

Durable agent loop

API → Workflow engine (Temporal-like)
         ├ step: research (LLM + tools) → save artifacts
         ├ step: draft (LLM) → create approval record
         ├ WAIT human decision (timer + reminders)
         ├ step: execute tools via outbox (idempotent keys)
         └ step: notify requester

Crash/deploy: resume from cursor; never re-send without idempotency

Deep Dive: Approval UX & Safety

Show diff of proposed CRM fields and email body with sources
Allow edit-then-approve; treat edits as new model-visible feedback
Separate draft identity from execute identity (least privilege)
Timeouts: auto-escalate to secondary approver, never auto-approve irreversible actions

Bottlenecks at Scale

Approver bottlenecks — load-balance queues by skill
Long-running waits — durable timers, not thread sleeps
Duplicate executes on retry — outbox + idempotency keys mandatory

Failure Modes & Monitoring

SLO	Target	Alert
Approval SLA	< 4 business h	Remind + escalate
Duplicate side effects	0	Page
Workflow success	> 95%	Warn
Stuck in WAIT	< 2% > 24h	Sweep job

AI & FDE Interview Room