On 21 October 2026, Rob Coward from the Errand AI team spoke at London DevOps AI #002: Spec-Driven, Guardrailed AI Coding, hosted at Civo Tech Junction in Shoreditch. London DevOps AI is the new identity of the long-running London DevOps meetup, refreshed after its first 100 events, and this second edition focused on how to get reliable, safe results from AI coding agents.
From Do-While to Loop-Until
Coding agents are brilliant at writing code and terrible at delivering features. The usual workflow is a
do-while loop: prompt the agent, let it build something, then try to reverse-engineer what it made and decide
whether it’s right — without ever seeing the options it considered and rejected along the way.
The talk proposed inverting that loop. Humans make the engineering trade-offs and write the specification; the agent then loops until the code matches the spec. Review moves up to the spec, where saying “no” is cheap, and the decisions about how a feature should work stay with the people accountable for it.
This is the same spec-driven approach we use to build Errand itself, with OpenSpec defining each change as a proposal, specs and tasks before an agent writes any code.
Also on the bill
The evening’s second talk came from Ace Vinayak, founder of Pluto Coach Ltd, with The Agent that Dropped the Table: Fail-Closed Guardrails for AI Coding Agents. It covered placing a deterministic policy gate in front of every tool call an agent makes, with allow/ask/block tiers and policies treated as code — a natural companion to spec-driven delivery.
Thanks to the London DevOps AI organisers, to Civo for the venue, and to the evening’s sponsors: Adaptavist, BIMP, Prism Digital and Tyme Technologies.
The slides from our talk are available here.
These are the links to the resources we mentioned during the talk: