Agent orchestration frameworks: what still breaks in production

Orchestration frameworks are good at drawing a path: retrieve, tool, generate, loop. Production is good at interrupting that path. A send needs a person. A CRM write needs a person. A customer replies three days later. The process that looked like a graph is now a wait. What still breaks is not “which of nine frameworks we killed in a lab.” It is waking up with the same state, the same thread, and the same authority model.

Agent Inbox plus Neuro OS Ask is a durable wait surface: the mailbox holds the conversation; the control plane holds the approval; the role resumes without pretending a chat tab was a queue.

Pause is easy; resume is not

Most runtimes can throw an interrupt or emit a “needs human” event. Fewer can guarantee that a week later the worker that resumes is looking at the same draft, the same tool permissions, and the same inbox. In-memory graphs die with the pod. Browser tabs die with the laptop. Redis keys expire. Someone approves a stale draft and you send last Tuesday’s price.

Durable resume needs: a workflow id, serialized state, a draft id that is not regenerated on retry, and an inbox thread id that is the customer’s reality. Email is helpful here because the other party already treats the thread as the source of truth. Your orchestrator should attach to that thread, not invent a parallel transcript.

Approval is a transaction

Ask is not a toast notification. It is a two-phase commit: draft prepared, human decision recorded, send or write executed once. If the orchestrator retries the whole graph, it must not create a second Ask card and a second send. Idempotency keys belong on draft creation and on send, the same way they belong on webhook handlers.

CRM writes use the same transaction shape. “Update the deal and email the buyer” is one review if you can show both payloads together; it is two if different owners must sign. Either way, neither happens because a model was confident.

Time is a first-class edge

Orchestrators often assume the next node runs now. Email introduces hours and days. A good production design has timers that draft a bump, then wait on Ask — they do not auto-nudge. It has SLA clocks that escalate by forwarding the thread into Neuro OS, not by paging a generic Slack webhook with a truncated body.

When the customer replies, the orchestrator is not “starting a new run from the original prompt.” It is continuing a thread. Load the mail. Load the last approved state. Do not replay the whole history of tool calls unless you enjoy paying for tokens and repeating side effects.

What frameworks still owe you

Pick LangGraph, ADK, Crew, Mastra, or anything else because of language and team skill — see nine frameworks and where email fits. Demand from all of them: externalizable state, human interrupts that survive deploys, and tools that can be scoped to one inbox. If a framework only offers “the model decides to call send,” wrap send yourself.

Do not keep mailbox passwords in orchestrator skills. Do not use a second agent as the approver. Do not claim you benchmarked the market into a winner. The production question is whether a person can find the waiting work, understand the thread, and finish the transaction.

Email plus Ask is not a framework. It is the wait primitive companies already understand. Orchestration is what you put around it so the role wakes up in the same place it went to sleep.

Agent Inbox gives each role a mailbox people can reply to, with forwarding into Neuro OS when a human must see the thread. Outbound mail defaults to Ask. Run the role on Neuro OS. To scope the first inbox, get started.

Explore Neuro OS →