Context Is the New Code (Patrick Debois, AI Engineer)

Patrick Debois (the DevOps coiner, now at Tessl) opens the architect track at the AI Engineer conference. The talk pitches the Context Development Lifecycle (CDLC) as the missing discipline around AI agent context — a direct callback to his 2009 DevOps move ("what if ops looked more like dev?" → "what if context is the code?").

Key claims

• Context is the new code, because it's being generated the same way code is. We're also turning code back into context (skills, workflows that previously required scripts can now be a few markdown lines + the model figures it out).
• The Context Development Lifecycle: Generate → Evaluate → Distribute → Observe. Infinity loop, DevOps-shaped. The talk's central artifact.
• Generate is the part everyone does. Prompts, instruction files (he's pleased a soft agents.md standard is emerging — "boo Claude for still calling it CLAUDE.md"), pulling lib docs, MCP-style pulls, spec-driven planning.
• Evaluate is the gap. Context changes have unknown impact and "YOLO, looks good to me" is the default. Layers of testing (lint → Grammarly-style readability → LLM as Judge → sandboxed agent tests), with error budgets, not pass/fail, because evals are non-deterministic.
• Distribute = the package/registry pattern. Skills are emerging as a package format all the major coding agents now recognize. Predicts dependency hell for context. Security: Snyk now scans context for credentials/third-party exposure. AI SBOM for provenance (which model built which skill).
• Observe = read agent logs. When an agent says "I'm missing X," if multiple devs hit the same gap, surface it and create org-wide context. PR feedback is implicitly feedback on the context that produced the PR. Production failures should write tests back into the context loop.
• Sandboxes don't filter context. Your agent loads agents.md and skills with no restriction — sandboxes restrict execution, not loading. He proposes a context filter — a web-application-firewall analog — that strips prompt injections / unsafe patterns before the agent reads them.
• Three nested loops as the coda: solo (you crafting markdown) → team (reflex to add context on every gap) → org-of-teams (fix once, every team benefits = the flywheel).
• "LLMs are just the engine. If you give the engine the wrong fuel, which is context, they're not going to perform." You can't tune the LLM, but you can engineer the context.

Tessl plug

One sentence at the end — Tessl implements pieces of the CDLC. No deep architecture details on stage. See Tessl.

Q&A worth keeping

• Audience question on "exotic" forms of context (architectural specs, consistency-as-eval — running the same prompt N times in parallel; if outputs are wildly different, the original spec was too loose). Patrick mostly punted but noted: "the piece people are underestimating is once you start writing context instead of code, you're going to spend that time writing the right evals."

Cross-source resonance

• Same conference, paired with Harness Engineering (Ryan Lopopolo, AI Engineer) — Patrick explicitly references "harness engineering" on stage as a parallel discipline. Ryan's talk is the operational counterpart (run the agent harness; this talk is the lifecycle around the context that goes into it).
• CDLC vs Context Engineering (IBM) — Patrick's CDLC is the process view (lifecycle phases). Martin Keen's framing is the architectural view (four pillars: connected access, knowledge layer, precision retrieval, runtime governance). Same field, complementary lenses.
• Skills as package format — directly parallels Claude Code's skill system and the Printing Press CLI factory. Both are concrete instances of Patrick's "packaging" phase.
• LLM as Judge as eval mechanism — also central to Ryan Lopopolo's reviewer-agent pattern. Two independent practitioners adopt the same primitive.
• Agent log standards (Agent ND) — anticipates the kind of observability layer described in Harness (LLM Agents) (Praveen's 5-component view).

Cross-links

• Patrick Debois · Tessl · AI Engineer · Context Development Lifecycle · LLM as Judge · Context Engineering · Harness (LLM Agents) · Harness Engineering (Ryan Lopopolo, AI Engineer)

Source

• Context Is the New Code — Patrick Debois, Tessl.md