Agent Runs That Survive the Crash
Every incumbent sells durability as a runtime migration — move your orchestration into their platform and rewrite your workers. BluePages ships the skill-native slice: checkpoint state, resume with idempotency guards, and fence duplicate workers — three pay-per-call skills your existing pipeline calls over x402.
Checkpoint → Resume → Fence
The three durability primitives, in the order your pipeline uses them — priced per call, paid in USDC via x402.
Snapshot state at every step boundary
Each checkpoint captures the pipeline position, an application state blob (content-addressed refs keep LLM-sized payloads out of the snapshot), accumulated x402 spend against the run budget, and per-step input hashes for replay validation. Checkpoints are hash-chained per run, so resume tooling detects a forked or tampered history before trusting it.
View schema, trust score, and live health →
Recover instead of restarting
Exchange a resume token for a validated resume plan: chain integrity verified, input-hash drift against the current pipeline definition detected, and each completed step assigned skip, verify, or rerun — with unprovable side-effect steps surfaced instead of silently skipped. Resume-count limits and per-resume spend deltas keep recovered runs auditable against one budget.
View schema, trust score, and live health →
Guarantee exactly one worker owns the run
Exclusive heartbeat-renewed execution leases with monotonic fencing numbers: a zombie worker holding a stale lease is rejected at the write, not trusted at the read. Missed-heartbeat expiry marks a run orphaned and eligible for takeover — the trigger that starts a resume — with lease transfer for deploy draining and priority waiter queues for contended runs.
View schema, trust score, and live health →
Three Failure Modes, One Recovery Contract
Long-running agent work fails in predictable ways. Durability turns each one from a full restart into a bounded, auditable recovery — and composes with spending limits, PipelineGuard breakers, and ComplianceKit audit trails you already run.
The 3 AM crash
A worker dies at step 7 of 9. Without checkpoints the run restarts from step 1 — re-paying every x402 fee and re-running every side effect. With a checkpointed frontier, the resume plan re-executes only what is unproven.
The duplicate execution
Two orchestrator replicas pick up the same queued run, or a presumed-dead worker comes back mid-recovery. Fencing numbers bind writes to the current lease holder, so the stale copy fails fast instead of double-spending.
The silent drift
The pipeline definition changed between crash and resume. Input-hash drift detection flags steps whose inputs no longer match what was checkpointed, so the resumed run never silently mixes two versions of the same pipeline.
Why Skills Beat a Workflow-Platform Migration
Adopt per pipeline
Durable-execution platforms ask you to move the orchestrator. Skills let you checkpoint one flaky pipeline today and leave the other nine untouched — no SDK rewrite, no runtime lock-in.
Spend-aware recovery
Checkpoints carry accumulated x402 spend, and resumes re-attach to the same budget. A recovered run stays inside the spending limit it started with — recovery never becomes a budget bypass.
Built for MCP Tasks
The MCP 2026-07-28 Tasks extension standardizes long-running, resumable tool work. Checkpoint/resume/lease is the state layer that story needs — available as skills before the spec even ships.
Building Durability Infrastructure?
Deterministic replay validators, side-effect journals, saga compensators — every incumbent in this category sells a migration. The skill-native slice is wide open. List yours while it is.