LangGraph 1.0: agents that survive a restart

The feature list of LangGraph 1.0, generally available since October 22 after powering production agents at Uber, LinkedIn, and Klarna for over a year, reads like database marketing rather than AI marketing. Durable state. Automatic recovery. Checkpointing. Pause-for-approval APIs. That is precisely the point: the gap between an agent demo and an agent in production is mostly the gap between a process and a workflow, and 1.0 is the workflow half arriving.

The headline capability is durable execution. Agent state persists automatically; a server restart or an interruption mid-task resumes exactly where it stopped instead of losing the run. Checkpointing extends the same idea to arbitrary save-and-resume points without custom database plumbing, which is what multi-day processes actually require. The human-in-the-loop support is first-class API rather than afterthought: execution pauses for review, modification, or approval, then continues with the human’s input in state. Underneath sits the graph execution model, deterministic edges where you have decided the flow, model-driven nodes where you have not.

The reason this note exists is the overnight research run. A desk agent that reads the day’s filings, queries internal databases, and drafts summaries is a multi-hour workflow touching rate-limited APIs and flaky connections. Without durable state, any failure at hour three restarts hour zero, which in practice means nobody ships the workflow at all. With checkpoints, failure costs one step. Durability converts agent reliability from a property of the model into a property of the runtime, which is where an engineering organization can actually manage it. The same primitive carries the long-horizon ambitions of models like K2 Thinking: a 300-tool-call trajectory without checkpointing is a 300-step single point of failure.

The human-in-the-loop API deserves a compliance reading. Pause-for-approval as a framework primitive means the four-eyes check that a regulated workflow requires is a node in the graph rather than a bolted-on email thread. The approval, the modification, the identity of the approver, all of it lands in the persisted state, which is to say in the audit trail. Anyone who has retrofitted approval gates onto a system that never expected them knows what being handed this for free is worth.

A year of production at that scale also explains the design priorities. Uber-scale workloads do not fail politely; they fail mid-run, at volume, on infrastructure someone is rebooting. Features like automatic recovery read as conveniences in a demo and as table stakes in that environment.

The 1.0 designation itself is the unglamorous half of the news. Breaking changes are now expensive for the maintainers, the prebuilt agents migrate out to the langchain package while the core API surface holds still. Framework churn has been a real tax on agent engineering, every quarter’s rewrite invalidating last quarter’s hardening. A stability contract is what lets a platform team commit. The selection pressure this archive keeps documenting, away from clever loops toward deterministic workflows with model-driven nodes wired to standard tool access, now has a stable artifact to build on.

The caveats are the usual ones for framework bets. Durable execution binds you to LangGraph’s persistence model, which is a real dependency to govern like any other vendor surface. The orchestration layer does nothing about the quality of what each node does, so retrieval, validation, and observability stay your problem; a graph of bad nodes fails durably. And one framework’s GA does not end the churn elsewhere in the stack. What it ends is the excuse that agent infrastructure is too unstable to ship on, which for at least one layer of the stack is no longer true.

LangGraph 1.0 makes agent runs durable, resumable, and pausable for approval: the boring database virtues, finally applied to the layer that needed them before any desk could run agents overnight.