← All speakers

Bio, Work & Ideas

Jonas Templestein

Conference affiliation: Iterate · 2026

On this page

Jonas Templestein is the founder and chief executive of Iterate and a co-founder and former chief technology officer of Monzo, the British digital bank. After helping build Monzo’s technology and customer experience, he turned to AI-assisted software creation and infrastructure for transparent, adaptable agents.

From early startups to Monzo

Templestein founded the conference-calling service Hipdial and co-founded Ark, a Y Combinator-backed people-search company. Early software projects included Node.js middleware for Facebook session authentication and tools for synchronizing objects across connected clients.

After working at Starling Bank, he co-founded Monzo, where he emphasized building secure financial infrastructure that engineers could continually adapt. He also created the merchant emoji in transaction notifications, a small but distinctive expression of the bank’s customer-focused design. Monzo’s 2022 annual report identifies him as a co-founder and chief technology officer.

He left Monzo on December 1, 2023, after approximately nine years, warning that financial success should not displace experimentation for customers. His subsequent venture, initially associated with the name Nustom and now known as Iterate, explored AI-assisted software development. One early experiment, Garple, sold short domain names through software built with an AI coding agent and backed by human engineers.

Templestein also created a tree-sitter grammar converter for llama.cpp, connecting existing programming-language grammars with constrained model output.

A more inspectable architecture for agents

At AI Engineer Europe 2026, Templestein and Iterate founding engineer Misha Kaletsky explored an event-sourced agent harness built around several concrete architectural commitments:

  • Durable event streams: Record model output, configuration changes, errors, and scheduled actions in an append-only history that can be inspected and replayed.
  • Separate reducers from side effects: Reconstruct an agent’s state before triggering external actions, preventing restarts from accidentally repeating earlier model requests.
  • Internet-native agent extensions: Give agents addressable URLs and let independently deployed processors contribute webhooks, safety checks, retrieval results, or sub-agent responses across machines and programming languages.
  • Programmable streams with operational safeguards: Configure workers through events while accounting for authentication, secret handling, rate limits, and circuit breakers that interrupt runaway loops.
  • Eventually consistent agent architecture: Allow external services a bounded opportunity to supply context or safety signals without making every extension a blocking dependency.

The workshop treated this system as an experimental prototype, not a production-ready platform; its public demonstration lacked authentication and exposed practical implementation failures alongside the architectural ideas.

Read the topics behind these talks

1 conference talk

References