AI agents are getting better at making plans. The less glamorous question is whether they can survive a failed API call, a redeploy or a human who takes three weeks to click “approve”. Restate co-founder Giselle van Dongen argues that this infrastructure problem is now the real frontier for teams putting agents into production.
Watch Desk analysis
What happened
In the supplied account of van Dongen’s AI Engineer podcast appearance, Restate is presented as a durable execution layer for long-running, stateful agent systems. Its journal records the steps an agent takes, allowing a failed web search or model call to be retried from the point of failure rather than restarting the whole job.
The demonstration connects a research agent to Slack. A planner creates subtopics, parallel agents investigate them and a writer assembles the result. Human approval can pause the process without consuming serverless execution time, while a later response resumes it from the stored suspension point.
Key findings
- Recovery instead of restart
Failed tool calls can be retried from journalled progress rather than sending the entire workflow back to the beginning. - Agents with durable memory
Restate’s “virtual object” model gives each session an identity, isolated state and handlers that can keep operating over time. - Steering while work is live
Relevant user messages can be signalled into a running agent, while changed instructions can cancel the current chain and start a new one. - Human approval that survives the calendar
A process can wait for a Slack decision through restarts and redeploys, then continue where it paused. - Concurrency without session collisions
The account says Restate queues competing executions so two messages do not overwrite the same agent state. - A claimed speed advantage
Van Dongen cites 45ms p99 latency for a 10-step workflow, though the supplied account does not establish how that result behaves under production-scale load.
Why it matters
The practical takeaway is refreshingly unglamorous: if an agent is expected to run for hours, call several tools, wait for people and remember what happened, prompt quality is only one part of the engineering job. Recovery, state, cancellation and flow control decide whether the system behaves like software or like a particularly confident intern with no backup plan.
For builders, the useful design question is whether these guarantees belong in an orchestration layer, an agent framework or the application itself. Restate’s pitch is that teams can make ordinary handlers durable without hand-building every retry and recovery path.
Our read
This is a persuasive explanation of why agent infrastructure is becoming a first-class product category. The strongest evidence is the concrete failure-and-recovery demo, not the AI branding around it. Treat the 45ms figure as a reported result, then test the model against your own workload, failure modes and concurrency before making it part of the production stack.
What to watch
- Whether the claimed latency holds with thousands of concurrent sessions.
- How cleanly the virtual-object model handles agents that need to change direction mid-run.
- Which guarantees are available in self-hosted, BYOC and managed deployments.
- Whether more agent platforms adopt durable execution as a default rather than an add-on.
Discussion spark: For production agents, which guarantee matters most to you: retryable steps, durable human approval, isolated session state or the ability to cancel and redirect a live run?
Sources and evidence
- Giselle van Dongen Says Agents Aren't the Hard Part – the Infrastructure Underneath Them Is – finance.biggo.com (14 September 2026, 17:12 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.