What is a context window in AI?

The maximum number of tokens a model can process in a single request. It covers everything the model can 'see' for that call: instructions, files, conversation history, and tool results combined. Anything outside the window doesn't exist for the model.

How big are context windows in 2026?

Frontier models typically run 200,000 tokens to 1 million, with some offering 2 million. A million tokens is roughly 750,000 words, several large codebases' worth of text.

Does a bigger context window solve the coordination problem?

No. A bigger window holds more of what you have, but a decision that lives only in a Slack thread nobody captured isn't in any window at any size. Capture and delivery are separate problems from capacity.

Context Window

The context window is the maximum text, in tokens, a model can consider in one request: its working memory. Anything outside it is invisible to the model.

The context window is a model's working memory for a single request: the maximum number of tokens it can consider at once. Tokens are word-pieces, roughly ¾ of an English word each. Everything competes for the same window: system instructions, your prompt, files, conversation history, tool results. What doesn't fit, the model never sees.

How big, in 2026

Frontier models typically offer 200,000-token to 1-million-token windows, and some reach 2 million. A million tokens is on the order of 750,000 words. That sounds like "just paste everything in" territory. It isn't.

Why "paste everything" fails

Cost. You pay per input token on every request. Filling a million-token window on each agent turn is expensive at exactly the moment agents are making more requests than humans ever did.
Relevance. Models attend less reliably to material buried in the middle of an enormous context. Ten relevant paragraphs outperform a thousand pages containing them. Selecting those paragraphs is the whole discipline of context engineering.
The hard limit. The window can only hold what was captured somewhere. A decision made in a meeting and never written down is outside every window at every size. Capacity doesn't fix decision drift; capture does.

What fills a coding agent's window

A single coding agent turn rarely starts empty. Before your request arrives, the window already holds the system prompt, tool definitions, and standing instruction files such as CLAUDE.md or AGENTS.md. Claude Code's documentation notes that each session begins with a fresh context window, and that these instruction files plus its auto memory notes are loaded at the start of every session. Then come the files the agent opens, command output, and every tool result. Long sessions fill up from the bottom, which is why agents get summarized or restarted mid-project, and why constraints mentioned early in a session can drop out later.

Window vs RAG

RAG doesn't enlarge the window. It decides what goes into it, picking passages by similarity. The window is the capacity; retrieval and MCP tools are two ways of choosing what spends it.

Window vs memory

The window is per-request and ephemeral. What an agent keeps across sessions is agent memory, a different layer with its own problems; persistent memory for AI coding agents covers how the two interact. Lockstep's job is upstream of both: keep the team's approved decisions captured and current, then feed the relevant few into the window via MCP, instead of hoping they fit somewhere in a million tokens.