Announcing Durable Coroutines for concurrent, fault-tolerant agents with millisecond overhead
- Spawn concurrent tools and subagents as lightweight coroutines within your agent process, each with its own durable steps.
- React to steering, approvals, and tool results, while in the background.
- Persist model calls, guardrail checks, and tool results across concurrent branches with low overhead (avg. 3ms), and recover after crashes.
- Reuse sandbox connections, warm clients, and caches across tools and subagents within the running process.
See the code and how the scheduler works 🧵
- Spawn concurrent tools and subagents as lightweight coroutines within your agent process, each with its own durable steps.
- React to steering, approvals, and tool results, while in the background.
- Persist model calls, guardrail checks, and tool results across concurrent branches with low overhead (avg. 3ms), and recover after crashes.
- Reuse sandbox connections, warm clients, and caches across tools and subagents within the running process.
See the code and how the scheduler works 🧵
4
2
36
1,945
Spawn concurrent tools, then use select to wait for a tool result or new steering input. Each tool can contain multiple durable steps and executes within the same process. After a crash, all executed steps are recovered in the same order, also on another process.
1
0
0
201
restate retweeted
Wrote my first X article, about something we built into @Restatedev earlier this year:
𝗗𝘂𝗿𝗮𝗯𝗹𝗲 𝗖𝗼𝗿𝗼𝘂𝘁𝗶𝗻𝗲𝘀 are an extension of durable execution with structured concurrency, but very fast (avg. 3ms overhead per step) on 3AZs.
It is great when you want to record many fine-grained steps in a larger task, like in a durable agent.
Say you dispatch a conversation turn as a durable function call (with queuing, limits, etc) and then within the turn you want to record inference, guardrails, tool calls with little overhead, as if it was a durable process.
x.com/StephanEwen/st…
𝗗𝘂𝗿𝗮𝗯𝗹𝗲 𝗖𝗼𝗿𝗼𝘂𝘁𝗶𝗻𝗲𝘀 are an extension of durable execution with structured concurrency, but very fast (avg. 3ms overhead per step) on 3AZs.
It is great when you want to record many fine-grained steps in a larger task, like in a durable agent.
Say you dispatch a conversation turn as a durable function call (with queuing, limits, etc) and then within the turn you want to record inference, guardrails, tool calls with little overhead, as if it was a durable process.
x.com/StephanEwen/st…
1
3
19
1,384
Replit Agent switched from Temporal to Restate for three main reasons
- Lower latency and cost efficiency:
Restate executes durable steps inline, adding just a few milliseconds of overhead per step. This allows Replit to make thousands of operations durable without the overhead as a separate activity.
- Unique building blocks for stateful, concurrent agents:
Virtual Objects and Durable Coroutines simplify coordinating long-lived agent sessions, parallel subagents, user steering, approvals, and cancellation, and make the architecture easier to extend.
- Restate BYOC:
Fully managed deployments in Replit's own cloud account keep sensitive data private and support both its multi-tenant cloud stack and dedicated enterprise offerings.
With Restate, Replit can use fast, durable execution inside the agent loop, giving them a simpler, more resilient architecture.
Replit now runs tens of Restate cells across regions and availability zones, and handled a major promotional traffic spike of nearly 25 000 durable actions per second per cell a few weeks after migration.
Read the full story, including how Replit built its agent architecture on Restate 🧵
- Lower latency and cost efficiency:
Restate executes durable steps inline, adding just a few milliseconds of overhead per step. This allows Replit to make thousands of operations durable without the overhead as a separate activity.
- Unique building blocks for stateful, concurrent agents:
Virtual Objects and Durable Coroutines simplify coordinating long-lived agent sessions, parallel subagents, user steering, approvals, and cancellation, and make the architecture easier to extend.
- Restate BYOC:
Fully managed deployments in Replit's own cloud account keep sensitive data private and support both its multi-tenant cloud stack and dedicated enterprise offerings.
With Restate, Replit can use fast, durable execution inside the agent loop, giving them a simpler, more resilient architecture.
Replit now runs tens of Restate cells across regions and availability zones, and handled a major promotional traffic spike of nearly 25 000 durable actions per second per cell a few weeks after migration.
Read the full story, including how Replit built its agent architecture on Restate 🧵
2
7
27
138,974
Learn how DOSS rebuilt their workflow engine for agents: with TypeScript workflows and Restate providing durable execution underneath.
Great work and deep dive, DOSS! 🚀
Great work and deep dive, DOSS! 🚀
What does it mean to be an AI-native platform?
We get that question a lot here. Perhaps one of the best examples is the way we handle our Workflows. When we first built DOSS, Workflows were done in a no-code visual editor, backed by our own language.
That worked great for a human to edit, but the people building workflows at DOSS increasingly aren't people. They're Agents. And Agents are dramatically better at reading and writing Typescript than clicking through a visual editor or hand-assembling a JSON graph.
So we rebuilt our Workflows to be the friendliest they could be to Agents, while maintaining the visual editor capability that some customers prefer.
Being AI-native means building a platform that's built as much for Agents as it is for people.
@kominko wrote about the change in one of our latest engineering blogs.
Read it here: doss.com/resources/eng-…
0
2
11
569
Restate v1.8.1 is out 🚀
🌊 Smooth hand-over when migrating Kafka topics to the new Kafka integration
🗑️ Remove several deployments at once
☸️ Kubernetes ConfigMap changes applied without restarts
🐛 Plus fixes, security updates and UI updates
github.com/restatedev/res…
🌊 Smooth hand-over when migrating Kafka topics to the new Kafka integration
🗑️ Remove several deployments at once
☸️ Kubernetes ConfigMap changes applied without restarts
🐛 Plus fixes, security updates and UI updates
github.com/restatedev/res…
1
2
16
936
restate retweeted
Also in 1.8: a CLI built for scripts and coding agents, with structured --json output, structured errors, and --dry-run / --yes. Plus SQL tables to inspect your vqueues and many more improvements.
Full announcement 👇
restate.dev/blog/announcin…
Full announcement 👇
restate.dev/blog/announcin…
0
1
4
157
restate retweeted
Restate v1.8 is out 🚀
The new virtual-queue scheduler is on by default, even for existing clusters.
"Why is this invocation still waiting?" Now you can see exactly why in the UI. 🔍
Tweak concurrency rules live to shape your traffic wrt your budget and needs!
The new virtual-queue scheduler is on by default, even for existing clusters.
"Why is this invocation still waiting?" Now you can see exactly why in the UI. 🔍
Tweak concurrency rules live to shape your traffic wrt your budget and needs!
1
6
11
454
restate retweeted
Sometimes it's the small things that make the big difference. P99 append latency in @restatedev 1.8: ~50ms → ~5ms 📉
2
5
27
1,165
restate retweeted
Congrats to @StephanEwen and the @restatedev team on their $20M Series A! We backed Restate at seed and are thrilled to keep building with Stephan, Till, Igal, Ahmed and the rest of the Restate team.
Software has always had to survive interruption, but AI agents raise the er, call models and APIs, use tools, and change direction mid-task. When something breaks, there is far more to preserve. It's no longer only the execution that needs to survive, but the state and communication around it.
Restate's founders created Apache Flink, which became foundational infrastructure for real-time data systems. With Restate, they built durable execution on a purpose-built distributed log, so execution, state, and communication live in one lightweight system with no external database required.
It's already running in production. Restate handles all durable orchestration for Replit Agent, and hundreds of companies, from leading AI companies to the Fortune 500, rely on it for mission-critical workloads.
We're proud to continue our partnership with Stephan and the Restate team as they make durability a property of the infrastructure itself. Let's go, team 🚀!
Software has always had to survive interruption, but AI agents raise the er, call models and APIs, use tools, and change direction mid-task. When something breaks, there is far more to preserve. It's no longer only the execution that needs to survive, but the state and communication around it.
Restate's founders created Apache Flink, which became foundational infrastructure for real-time data systems. With Restate, they built durable execution on a purpose-built distributed log, so execution, state, and communication live in one lightweight system with no external database required.
It's already running in production. Restate handles all durable orchestration for Replit Agent, and hundreds of companies, from leading AI companies to the Fortune 500, rely on it for mission-critical workloads.
We're proud to continue our partnership with Stephan and the Restate team as they make durability a property of the infrastructure itself. Let's go, team 🚀!
6
10
46
2,081
restate retweeted
Founded by Apache Flink veterans, the startup is taking on workflow heavyweight Temporal.
spr.ly/6016BGs9Vn
spr.ly/6016BGs9Vn
4
14
87
47,632
restate retweeted
We’ve raised a $20M Series A led by Singular with participation from Redpoint Ventures and Capital One Ventures.
Restate is already powering Replit Agent, enterprise workflows at DOSS, and critical applications at Fortune 500 institutions.
Now we’re doubling down on making r, more efficient and easier to operate at scale for the next generation of applications and AI agents.
👇 More on what we’re building and where we’re going next
Restate is already powering Replit Agent, enterprise workflows at DOSS, and critical applications at Fortune 500 institutions.
Now we’re doubling down on making r, more efficient and easier to operate at scale for the next generation of applications and AI agents.
👇 More on what we’re building and where we’re going next
14
11
90
8,449
We’ve raised a $20M Series A led by Singular with participation from Redpoint Ventures and Capital One Ventures.
Restate is already powering Replit Agent, enterprise workflows at DOSS, and critical applications at Fortune 500 institutions.
Now we’re doubling down on making r, more efficient and easier to operate at scale for the next generation of applications and AI agents.
👇 More on what we’re building and where we’re going next
Restate is already powering Replit Agent, enterprise workflows at DOSS, and critical applications at Fortune 500 institutions.
Now we’re doubling down on making r, more efficient and easier to operate at scale for the next generation of applications and AI agents.
👇 More on what we’re building and where we’re going next
14
11
90
8,449
restate retweeted
Three years ago, we started Restate because we felt that building reliable distributed applications was just too hard. We were solving the same problems around failures, recovery, and consistency again and again.
Today, I'm excited to share that we've raised a $20M Series A led ipation from Redpoint Ventures and Capital One Ventures.
The need for solutions like Durable Execution has only grown since we started. Today, the majority of new Restate use cases are AI agents: long-running processes, coordinating across many tools and APIs, that need to keep working correctly through all sorts of failures and interruptions, coordinate sub-agents, or schedule work.
At the same time, our customers have pushed Restate into places that expanded our own view of what Durable Execution could become. Replit moved the execution of Replit Agent onto Restate. DOSS runs high volumes of enterprise workflows on Restate and Fortune 500 institutions are using Restate for complex transactional workflows.
We think this is just the beginning and excited to build what we believe will be a fundamental pillar of all infrastructure in the future.
Read more here: restate.dev/blog/announcin…
Today, I'm excited to share that we've raised a $20M Series A led ipation from Redpoint Ventures and Capital One Ventures.
The need for solutions like Durable Execution has only grown since we started. Today, the majority of new Restate use cases are AI agents: long-running processes, coordinating across many tools and APIs, that need to keep working correctly through all sorts of failures and interruptions, coordinate sub-agents, or schedule work.
At the same time, our customers have pushed Restate into places that expanded our own view of what Durable Execution could become. Replit moved the execution of Replit Agent onto Restate. DOSS runs high volumes of enterprise workflows on Restate and Fortune 500 institutions are using Restate for complex transactional workflows.
We think this is just the beginning and excited to build what we believe will be a fundamental pillar of all infrastructure in the future.
Read more here: restate.dev/blog/announcin…
70
34
353
465,766
🧵Did you see the new project experience in Claude? Want to build something like it for your own agent?
These same primitives work for coding agents, research agents, or whatever you’re building next. ⬇️
These same primitives work for coding agents, research agents, or whatever you’re building next. ⬇️
2
0
13
1,257
4. Unified overview
Build your dashboard over durable project and agent state: progress, pending approvals, and results.
Build your dashboard over durable project and agent state: progress, pending approvals, and results.
1
0
1
93