Context is the new code: the development lifecycle for AI agents

What's in this session

Agents are increasingly driven by context: prompts, skills, rules, docs, specs, Slack history, and the instructions teams package around their work. If that context determines what agents produce, it cannot stay outside the engineering system.

Patrick Debois maps the software development lifecycle onto a Context Development Lifecycle: lint markdown skills, test prompts with LLM-as-a-judge evals, package context into registries, scan it before release, observe agent logs and production feedback, and feed what you learn back into better context.

For platform and developer-experience leaders, the takeaway is that context engineering becomes a team capability. Mature agent organizations will measure retries and hops, manage context drift like technical debt, and create multiplayer systems where good context compounds across the company.

Inside the recording

1. 00:00 Context is the new code
   Patrick maps software lifecycle practices onto the context that now drives agents.
2. 03:00 Linting and testing agent context
   Skills, prompts, and context files can be checked for structure, prose, and expected output.
3. 06:00 Registries for blessed organizational context
   Public skills are not enough; teams need private, tested context packages they can trust.
4. 09:00 The human harness
   Better documentation, feedback, and practice help humans create stronger agent context.
5. 12:00 Measuring quality and managing drift
   Retries, output quality, and context drift become the signals to maintain like technical debt.

More sessions on agent infrastructure

• Building Minions: agents on a 30-million-line codebase — Alistair Gray, Stripe
• Building a company-internal background agent system — Cole Murray, Open Inspect
• From Assisted to Delegated: Cloudflare's AI Engineering Stack — Rajesh Bhatia, Cloudflare