An agent can only use
the context it receives.

The prompt asks. The full context equips.

A prompt is the instruction: what to do, under which constraints, and in what form. The model’s context is the material it receives for a particular call. That may include the prompt, conversation history, tool definitions and results, retrieved documents, and a summary of earlier work.

“Context engineering” is a useful name for the work of curating that changing input. It shifts the design question from “What wording should we add?” to “What information should be available for this decision, and how will it stay current?”

The distinction matters most in multi-step systems. A request can be clear while the agent is missing the current policy, the result of its last search, or the reason an earlier route was rejected.

Bring together five things the next decision depends on.

Not every call needs every layer. A one-shot summary might need a source and a requested format. An agent deciding whether to run another check also needs the last result, the remaining uncertainty, and a clear stop condition.

Search the collection; pass forward the useful evidence.

Retrieval can find material outside the model’s current input. Context design decides which results deserve space in it. Returning every matching document or replaying an entire conversation may increase noise without improving the decision.

• Ask a specific question that matches the next action or decision.
• Keep source names, dates, and section references attached to relevant passages.
• Filter by the user’s permissions before exposing retrieved content to the model.
• Check whether a newer source supersedes an older one; keep genuine conflicts visible.
• Load more evidence when the first set is incomplete, instead of guessing from a thin result.

This is related to, but broader than, RAG. Retrieval is one way to add evidence. The agent also needs instructions, tool results, and a reliable account of its own progress. See where retrieval and fine-tuning fit ↗

Carry forward the work, not every line of the transcript.

Long-running agents accumulate messages, tool output, retries, and details that later become outdated. When the full history no longer helps with the next step, make a compact handoff that preserves the current goal, established facts, decisions, source references, failed attempts, and open questions.

Goal for this step: [the next decision or action]
Current facts: [fact] — [source and date]
Decisions already made: [decision and reason]
Actions and results: [what was tried; what happened]
Still unknown: [question, conflict, or missing source]
Available tools: [what this run may do]
Stop or ask when: [condition]

This is a handoff shape, not a required format. Keep durable memory separate from the working set assembled for each call. When using an API’s compaction feature, follow that provider’s state and replay rules; a compacted context is not a substitute for retaining the sources behind important claims.

Test what the agent receives, not only what it answers.

A long context window does not by itself show that a model can find and apply every relevant detail. In a 2024 study, researchers found that several evaluated models performed worse on some long-context tasks when the relevant passage appeared in the middle rather than near an edge. The result describes the models and tests in that paper; current systems and real workflows need their own evaluation. For the mechanics of tokenization and context limits, read the guide to tokens and context windows ↗.

Build a small test set from real decisions and vary the conditions around them:

Track task success, evidence quality, incorrect or unsupported claims, tool choices, review effort, context size, latency, and cost. The right measures depend on the task. A good answer on one short example does not establish dependable performance across longer runs.

For a focused test plan across context lengths, evidence positions, languages, and missing information, see how to evaluate a long context window ↗.

Context can guide a tool call. It cannot authorize one.

Documents, web pages, and tool results may contain hostile or misleading instructions. Treat retrieved content as data with provenance, and enforce access and approval in the tools and application. A paragraph in the model’s context does not replace a permission check.

For each action, expose only the required capability, validate its arguments, and check the caller’s authority at the point the action runs. If the action has a consequential side effect, pause at the human review boundary your product requires.

Context connects the parts of an agent.

Read what makes a system an AI agent ↗, how memory is selected and updated ↗, and why agents need evidence, not confidence ↗.