Mateusz Jagodziński

Your AI agent runs for days on Temporal. It retries, waits, and resumes across failures — reliably. But when it's about to do something irreversible, who approves it, and on what surface?

Durable execution engines make agents survive; they don't give a human anything to see or sign off on. That's an interface problem, not an engine problem — and it's where agentic projects die (Gartner expects over 40% to be scrapped by 2027, mostly over governance and control).

This talk shows how to build that missing layer: human-in-the-loop for long-running agents on Temporal. An approval gate is a signal the workflow waits on; a timeout is a timer; a human can edit state before the run resumes — all rendered on the same graph the engine actually executes, because a person making the decision needs workflow context, not a distributed trace. The layer is engine-agnostic by design: Temporal is the reference adapter, with a port for other engines.

You'll get the architecture, the signal/timer plumbing, a production case where an agentic-AI platform integrated this UI layer in about a day, and a live demo — including what broke: race conditions on signals, lost approvals on replay, approval-gate UX (who signs, what they see, with the EU AI Act in the background), and visualization–code drift.

Takeaways:
- Human-in-the-loop maps cleanly onto durable-execution primitives (signals, timers) — accountability becomes part of the workflow, not a bolt-on.
- The surface a human approves on must show workflow context, not engine telemetry.
- Reliability comes from the engine; accountability comes from the layer on top.

Mateusz Jagodziński is a Solutions Engineer at Synergy Codes, a 70-person software company specializing in diagramming and workflow tooling, and the team behind Workflow Builder. He has spent seven years there working in React, specializing in data visualization solutions and interactive diagramming interfaces.

He is usually the first technical person a new customer talks to: working out the requirements, picking the technologies, and then shipping the integration. Recently that has meant building the human layer on top of agentic AI platforms - the surface where a person sees what an autonomous agent is about to do and approves it before it happens - alongside the teams running those systems in production.

He has spoken at React Summit and React Summit US on extendable React architecture and design-token-driven frontend development.