Context is not given to an agent. Context is curated for it.
From the Daily trail
Agent Runtimes Are Moving State Out of the Shadows: Shows runtime state, queues, hooks, and delivery paths becoming visible.
Agent Context Becomes Runtime State: Daily evidence for context becoming an explicit runtime payload.
Perception is a product decision
The easiest way to misunderstand agents is to imagine the model staring directly at the whole world. It never does. The world is assembled into a view: prompts, recent turns, summaries, file slices, tool descriptions, repository metadata, permissions, and maybe a compacted memory of what happened before.
That makes context a newsroom problem and an operating-system problem at once. What is included becomes part of the agent's reality. What is excluded becomes invisible unless a tool can fetch it later.
The context layer is where magic becomes accounting
MCP's official specification names resources, prompts, tools, roots, sampling, and elicitation as protocol features. OpenAI's conversation-state docs describe several ways to carry context across turns. The Git Reporter's daily beat has seen the same pressure in repos: context becomes runtime state, summaries become infrastructure, and file tools turn into sensors.
This is the chapter where readers should feel the awe of constraint. The model's mind is vast, but the agent's working world is made from finite packets. Great agent systems do not merely add more tokens; they choose what deserves to become part of the next action.
What the agent cannot see matters too
Blind spots explain many failures that look like stupidity. The agent did not see the hidden file. The subagent did not inherit the policy. The compacted summary preserved the plot but lost a detail. The tool schema was vague. The UI told the human one thing while the runtime saw another.
A serious reader should finish this chapter with a new question for every agent product: show me the viewport. What evidence, memory, tools, and boundaries are actually in the room when the model decides?