Research · Area 03 of 07
Orchestration & Durable Execution
An agentic workflow that cannot recover from a crash, deploy, or interrupted approval is not ready for consequential authority, and reliability at that scale is an infrastructure discipline, not a prompt problem.
Stoa Labs studies how long-running and multi-agent work stays correct: durable state, safe and idempotent side effects, resumability, compensation, containment, and the verification of runs that span hours or days.
We do not build a competing durable-execution engine; the open problems sit one layer up, in correctness, idempotency, and verification, and that is where our work concentrates.
01 · Definition
Two halves that converged
This area covers coordinating multiple agents, and making agents reliable enough to run for minutes, hours, or days across crashes, deploys, and human pauses. The two halves converged in 2026: orchestration frameworks added persistence, durable-execution engines added agent-native primitives, and they now meet in the middle.
Five sub-areas with settled terminology:
- Multi-agent orchestration. How agents coordinate. The production default topology in 2026 is supervisor (orchestrator-worker) plus handoffs; teams, swarms, and graph pipelines cover the rest.
- Durable execution. Workflow runtimes that journal every step so a function resumes exactly where it stopped after a crash, applied to agents so a redeployed agent does not repeat tool calls, payments, or approvals.
- Durable long-running agents. Checkpointing, replay and resumability, time travel to prior checkpoints, and human-in-the-loop pauses that persist indefinitely.
- Background and scheduled agents. Agents running detached in cloud sandboxes, triggered by schedules, APIs, or repository events.
- Agent-to-agent interop. The settled two-layer stack: MCP for vertical tool access, A2A for horizontal agent coordination, both under Linux Foundation governance.
02 · Why it matters
From sessions to systems
Agents are graduating from sessions to systems: long-running, multi-step work with real side effects. Reliability at that scale is an infrastructure discipline, not a prompt problem, and it is the layer where enterprise deployments currently fail.
As of early 2026 only around 11 to 14 percent of enterprise agent pilots reached production at scale, with coordination, observability, and governance (not model quality) as the cited blockers. Meanwhile the raw async coding agent is commoditizing, which pushes value up into orchestration, reliability, and verification.
03 · Open questions
What we are trying to find out
Exploration-mode questions: pursued to learn something, allowed to not pan out.
- Idempotency and exactly-once side effects. Durable execution gives resume, not rollback: compensation and idempotency are still hand-rolled per tool. Can they be made automatic across MCP tools and A2A calls?
- Versioning of in-flight agent state. A paused workflow that resumes after an upgrade "resumes into a different universe," and for agents the prompt and model are versions too. Safe migration of paused agent state is unsolved.
- Checkpoint granularity and replay drift. Where to checkpoint, how to handle an LLM’s mid-reasoning state, and how to replay non-deterministic model calls without divergence.
- Verification of long-running runs. How do you establish that an agent did the right thing across a three-day, multi-agent run?
04 · Directions
What this work could become
Candidate development directions, not commitments or year-one programs.
- Idempotency and compensation patterns for agent tool calls, packaged as framework-agnostic libraries and reference implementations. The most acute, least-solved reliability pain.
- Integration and migration expertise across the durable-engine and agent-framework seam (Temporal, Restate, DBOS with LangGraph-class frameworks): opinionated reference architectures and reusable adapters.
- Verification harnesses for long-running and multi-agent runs: trace-level correctness checks, replay-based regression, and approval audit trails. This connects directly to the Harness Engineering pillar’s test-harness work.
05 · Boundaries
What we deliberately avoid
Building a competing engine. The engine layer is crowded and heavily funded (Temporal raised $300M at a $5B valuation in February 2026 and pivoted to agents; Cloudflare and the hyperscalers ship runtime primitives directly). The opportunity is one layer up: correctness, idempotency, verification, and interop glue.
06 · Landscape · 2026-07
The landscape, July 2026
Temporal, Restate, DBOS, and Inngest differentiate mainly by deployment model (managed cluster, HTTP sidecar, Postgres-as-backend, serverless events). A new agent-native durable runtime category emerged with Cloudflare’s Project Think (April 2026), where each agent is a durable actor with checkpointable function calls and zero-compute hibernation, reframing request-response enterprise runtimes as a prior generation.
Scheduled and event-triggered cloud agents went mainstream (Claude Code remote tasks and routines, Codex Cloud, Cursor cloud agents), and Devin’s price drop signaled commoditization of the raw async agent.
Interop consolidated: A2A passed 150 organizations at its one-year mark with production use in financial services, insurance, and supply chain, and MCP plus A2A under one governance roof is the settled stack. Open gaps: discovery at scale, fine-grained authorization, and cross-protocol bridges.
Landscape notes are dated on purpose. This layer moves monthly, and we treat any claim older than a quarter as due for reverification.
07 · Sources
Selected sources
- Cloudflare, "Project Think" (April 2026); Temporal Series D and Replay 2026
- "Agent workflows are rediscovering durable execution" (May 2026)
- Linux Foundation, A2A one-year milestone (April 2026)
- Durable-execution engine comparisons (Temporal, Restate, DBOS, 2026)
Where this work goes
This research supplies the methods behind our harness and orchestration engineering services.