Every long-running agent pipeline has the same failure story. Twelve steps in, forty minutes of work done, $0.38 of x402 fees paid — and the worker gets OOM-killed, the deploy drains the pod, or the orchestrator process just dies. The run restarts from step one. Every completed step re-executes. Every paid skill charges again. And if two replicas both pick up the "failed" run, you get the duplicate-execution special: the same side effect landing twice, from two workers that each believe they own the run.
The industry noticed. This quarter, durable execution went from a niche workflow-engine concept to the loudest infrastructure category in the agent stack: Temporal built its entire Replay 2026 conference around durable agents, Restate, Inngest, Trigger.dev, and DBOS all repositioned as "the durable layer beneath agent SDKs," and Microsoft updated Durable Task specifically for AI agents. The pattern is validated. The problem is what adopting it costs: every one of those platforms is a runtime migration. Your pipeline moves into their workflow engine, their SDK wraps your steps, their infrastructure hosts your state.
The Skill-Native Slice
Today DurableRun.dev joins BluePages with three skills that deliver the durable-execution contract as pay-per-call primitives — the slice the workflow platforms don't sell, for pipelines that want recoverability without a re-platforming project.
Run Checkpointer ($0.001/call) snapshots pipeline state at step boundaries: position, retry counts, an application state blob (content-addressed external references for LLM-sized payloads), accumulated x402 spend, and input hashes of completed steps. Checkpoints are hash-chained per run — every snapshot references its parent, so resume tooling can detect a forked or tampered history before trusting it.
Crash Resume Orchestrator ($0.003/call) turns a checkpoint into a resume plan with idempotency guards. It validates the chain, replays step input hashes against the current pipeline definition to catch drift (a changed schema since the checkpoint invalidates blind resume), and classifies each completed step as safe-to-skip, must-verify, or must-rerun. Side-effect steps without idempotency keys are surfaced explicitly rather than silently skipped — the plan is honest about what it cannot prove already happened.
Execution Lease Manager ($0.001/call) guarantees exactly one worker owns a run at a time. Acquire returns a fenced lease token with a monotonic fencing number; checkpoint writes and downstream skills can require the current number, so a zombie worker holding a stale lease is rejected at the write, not trusted at the read. Missed heartbeats mark a run orphaned — and orphan takeover is precisely the trigger that starts a resume.
What This Costs
A pipeline running 200 multi-step runs per day, checkpointing at 6 step boundaries per run, with lease acquire/heartbeat/release per run and (say) 4 crash resumes a day:
- 1,200 checkpoints × $0.001 = $1.20
- ~600 lease operations × $0.001 = $0.60
- 4 resumes × $0.003 = $0.012
Call it $1.81/day for runs that survive crashes. Compare that to the alternative you're already paying: every restarted run re-pays every completed paid step. One recovered 12-step run with $0.38 of fees pays for ten days of the entire durability layer.
The Composability Loop
Durability compounds with the guardrails already on the marketplace. Checkpoints record accumulated spend, so a resumed run re-attaches to its spending limit budget instead of starting fresh — recovery can't be used to sidestep a cap. The resume plan's skip/verify/rerun verdicts compose with PipelineGuard circuit breakers (don't resume into an open breaker) and QueuePilot retry policies (resume is what happens after retries are exhausted, not instead of them). Lease audit logs feed ComplianceKit audit trails, and a resumed run's lineage — which worker, which checkpoint, how much was spent across attempts — is exactly the evidence an AgentSure incident claim or a VerdictLayer SLA dispute asks for.
And when the MCP 2026-07-28 release lands with its Tasks extension for long-running work, checkpoint/resume is the natural backing store for a task that outlives its connection. That's not a coincidence; it's the direction the whole stack is moving.
Restart-from-zero was fine when agent runs took eight seconds. Your runs take forty minutes now. Checkpoint them.