You can swap in the best model on the leaderboard and your agent will still fail - because the bug was never in the model. It was in what you fed it. This talk is a postmortem on the failures that actually block agents from shipping: the ones that happen before the first token is generated, in the invisible step where you decide what the model gets to see. After running multi-agent workflows across 1M+ executions a month, I can tell you the model is almost never the problem. The context is.

Prompt engineering taught us to obsess over wording. But in production, agents don't fail because a prompt was phrased poorly - they fail because the retrieval missed, the memory was stale, the tool schemas ate half the window, or the supervisor's summary dropped the one fact the worker agent needed. These are not prompting bugs. They are context engineering bugs, and they belong to a discipline most teams are doing by accident.

I'll make that discipline explicit by treating context assembly as a compile step - a pipeline that turns messy world state into the exact tokens the model sees:

- Context as a compiler, not a prompt - retrieval, reranking, memory injection, and tool-state snapshots, and why order changes the output
- Multi-agent context routing - what a supervisor hands to a specialist, what comes back, and what silently gets lost in translation
- Memory tiers under load - hot (recent turns), warm (session summaries), cold (long-term RAG) - and which one breaks first
- Compression that lies - when summarization saves your token budget and when it quietly poisons the next decision
- Token economics - the budget war between system prompts, tool schemas, and the context you actually need
- Debugging what the model saw - traces, diffs, and the one question that finds most agent bugs: "What was actually in the window?"

You'll leave with a mental model for context engineering as infrastructure, a checklist of failure modes to design against before they reach production, and a walkthrough of real traces where the fix was never a better prompt - it was better context.

Prompt engineering got us to the demo. Context engineering is what gets you to the millionth run.