Blog

Building production AI agents

Field notes on the hard parts of running agents for real: durable execution, crash recovery, human approval, and putting a cost on every step.

July 19, 20268 min read

Durable execution for AI agents: how checkpointing survives crashes

A crash shouldn't cost you a half-finished agent run. How checkpointing lets an agent resume from exactly where it stopped — same state, same position in the graph.

July 18, 202610 min read

Temporal vs LangGraph for AI agents (and when you need neither)

They aren't really competitors: one is a durability engine, the other an agent-authoring framework. Which you need depends on whether your gap is durability or agent structure.

July 17, 20268 min read

How to track the cost of every AI agent step

A single provider bill tells you nothing about which agent, run, or step spent the money. How to attribute cost at node granularity — and cap it before it surprises you.

July 16, 20268 min read

How to add human approval to an AI agent

High-stakes agent actions need a human to sign off. How an approval gate pauses a run as a durable checkpoint — for minutes or days — without holding a process open.

July 15, 20267 min read

How to resume an AI agent after a crash

A crash shouldn't cost you the whole run. How checkpointing lets an agent restart from exactly where it stopped, re-running only what was actually in flight.