Measured. Reproducible.
Honestly caveated.

Moving parts at runtime

The number of things that must be deployed, monitored, patched, and paged about, for a durable agent with human approval. Sourced from each project's own documentation (July 2026).

Sources: openai-agents docs (HITL & running-agents), langchain-ai LangGraph platform docs, inngest/agent-kit README, temporalio docs. Capability differences, not quality judgments. Temporal's durability is superb if you can run Temporal.

Code to get a durable, approvable, budget-capped agent

OpenAI's own canonical human-in-the-loop examples are 116–137 lines, and their "persistence" is a local JSON file with resume plumbing left to you. Here's the whole thing in Nexus, durability included:

// 1. define tool with needsApproval          (their SDK: ✓)
// 2. run, detect result.interruptions        (their SDK: ✓)
// 3. serialize RunState to a string          (their SDK: ✓)
// 4. design a DB schema for pending states   (you: build it)
// 5. build an approval UI / notification     (you: build it)
// 6. build the resume endpoint/worker        (you: build it)
// 7. handle crash mid-turn                   (you: can't, turn is lost)
// 8. token/USD budget caps                   (you: build it)
// 9. audit journal of every step             (you: build it)
// 10. replay for regression testing          (you: build it)

runtime.register(defineAgent({
  name: 'support', provider, model,
  tools: [lookup, refund],       // refund: requiresApproval
  budget: { maxUsd: 0.50 },      // hard cap, pre-flight
  approvalPolicy: ({ arguments: a }) =>
    a.amountUsd < 50 ? 'approve' : 'ask',
}));

const run = await runtime.run('support', task);
// paused? notify Slack (1 line), approve from anywhere:
await runtime.approve(run.id);
// crashed? resume. finished? replay. all built in.