Agents in production: ReAct loops, timeouts, and human-in-the-loop

Agent demos loop forever. Production agents need budgets, idempotent tools, escalation paths, and explicit stop conditions.

Aug 7, 202614 minPattern inspired by agent platforms (Uber-class / Stripe-class harnesses)

agents
react
hitl
reliability

Framed from public engineering talks, blogs, and OSS patterns. Not confidential internals or invented quotes.

A ReAct agent that “keeps trying” looks smart in a blog GIF and catastrophic in prod: runaway tool calls, duplicate side effects, and spend spikes. Public patterns from agent platforms and coding harnesses emphasize budgets, verification, and human gates.

Core loop with guardrails

flowchart TD
  Start[User goal] --> Plan[Think / plan]
  Plan --> Act[Tool call]
  Act --> Obs[Observe result]
  Obs --> Check{Budget / stop?}
  Check -->|continue| Plan
  Check -->|need human| HITL[Human approval]
  Check -->|done / fail| End[Final response]
  HITL --> Act

Hard budgets (ship all four):

Max steps / tool calls per run
Wall-clock timeout
Token / $ ceiling
Idempotency keys on mutating tools

Tool design for agents

Tools are the real API surface:

Narrow, documented side effects
Structured errors (retryable, fatal, needs_auth)
Sandboxing for code execution
Authz checked on every call with user/tenant identity

sequenceDiagram
  participant A as Agent
  participant G as Gateway
  participant T as Tool
  participant H as Human
  A->>G: tool_call(delete_resource)
  G->>G: policy check
  G->>H: approve?
  H-->>G: approve
  G->>T: execute + idempotency-key
  T-->>A: result

Planning vs reactive

Style	Strength	Weakness
Pure ReAct	Flexible	Wander / loops
Plan-then-execute	Auditable	Brittle plans
Workflow graph (LangGraph-style)	Explicit states	More upfront design
Swarm / multi-agent	Specialization	Coordination tax

Prefer explicit graphs for business-critical flows; keep free-form ReAct for exploratory internal tools with tight budgets.

Human-in-the-loop (HITL)

Insert HITL when:

Irreversible or high-cost actions
Low model confidence / policy uncertainty
Legal or brand risk

Make HITL a first-class state with SLA — not a Slack DM hope.

Failure modes

Infinite clarification loops with the user
Partial writes without compensating transactions
Prompt injection via tool outputs (treat as untrusted text)
Hidden fan-out: one user click → N agents → M tools

Eval trajectories

Unit-test tools; integration-test trajectories (golden paths + adversarial). Score: task success, step count, policy violations, human escalations.

What to ship for v1

ReAct (or graph) with max-step and timeout
Idempotent mutating tools
HITL for a defined action class
Trace every thought/tool/result (redacted)
Kill switch per agent type

Design interview punchline

“An agent is a state machine with an LLM as a fuzzy transition function — productionize the state machine first.”

The demo vs the pager