Context Engineering Explained: How to Give AI Agents the Right Information
By Daniel Mercer• 10 mins read
Published:
, Updated:
Learn how context engineering helps AI agents select sources, separate instructions from evidence, manage memory and stay accurate during longer tasks.
An AI agent can have the right model and a clear task, yet still work from the wrong information. It may quote an outdated draft, overlook a decision buried in a long conversation or treat a search snippet as if it were the full source. Improving the wording of the request will not necessarily fix those failures.
Context engineering is the work of deciding what information an AI system receives, how that information is organized and what remains available as the task develops. This guide explains the practical choices through an illustrative product-launch FAQ: a current brief, older planning documents and fresh source material that must be reconciled into one accurate result.
Context Engineering Goes Beyond the Prompt
A prompt tells the assistant what you want. The surrounding context supplies the material it can use to respond: instructions, conversation history, retrieved documents, tool descriptions and tool results. Context engineering considers that whole working set, including what should enter it later rather than being supplied immediately.
Anthropic describes the discipline in its engineering guide as an ongoing process of curating useful information for an agent. Its central point is relevance: the goal is enough high-value context to support the task, rather than the largest possible collection of text.
The company’s context-window documentation distinguishes the model’s active working context from the material used to train it. Instructions, messages, tool definitions and results consume space in an API request, and generated output also counts toward the window. A large capacity does not mean every supplied detail will be used equally well.
For our FAQ, the instruction might be to explain the product’s launch features for existing customers. The evidence includes the approved feature list and current availability details. The old positioning document may help explain a historical decision, but it should not silently define what customers can use today.
Why a Bigger Context Window Is Not the Whole Answer
Capacity and performance are different measurements. A model may accept a long document collection without reliably selecting the right fact from it. Adding more material can also introduce contradictions, irrelevant background and repeated versions of the same claim.
The 2023 Lost in the Middle paper examined multi-document question answering and key-value retrieval. It found that performance in the studied models depended on where relevant information appeared, with important details in the middle of long inputs often harder to use. This is evidence from those experiments, not a ranking of today’s models or a universal rule about every task.
The practical response is to test how your workflow uses information, rather than assume a capacity figure settles the question. For the launch FAQ, place a crucial availability restriction in different locations within a controlled document set. Check whether the final answer preserves it each time, including when plausible but obsolete material is nearby.
A longer input can still be appropriate when the task requires broad comparison or detailed analysis. Cutting a document too aggressively may remove the qualification that changes its meaning. Context engineering is therefore selective organization, not a contest to make every request as short as possible.
Build a Clear Brief Before Adding Documents
Start with the decision or deliverable the agent must produce. In our example, the output is a customer-facing FAQ covering the launch features, eligibility and known limitations. It should identify unanswered questions and remain a draft until the product owner reviews it.
Next, identify which source governs which kind of claim. The approved product brief defines the launch scope. A current release record establishes what has actually shipped. A support note may reveal a recurring customer question, but it does not authorize a new feature promise.
This is an editorial policy for the worked example, not a universal ranking of document types. Another organization may use a signed specification or a maintained product database as its authority. What matters is that the agent is told how to resolve disagreements rather than being left to infer authority from a filename or confident wording.
Give sources useful labels and retain their identity. A document title, revision date, owner and link help explain where a claim came from. Mark a superseded plan as historical and name the document that replaces it. Newer is not automatically more authoritative if the new file is an unapproved draft.
A compact brief for this example could contain five parts:
The audience and the questions the FAQ must answer
The approved launch scope and the source that defines it
The current release evidence and when it was checked
Known conflicts, missing facts and claims that must remain unresolved
The required output format and the limit to drafting rather than publishing
Ask the assistant to identify the sources it needs before loading an entire archive. If it cannot find current eligibility information, that is a gap to resolve. Filling the gap with a reasonable-sounding inference would make the answer more complete in appearance while making it less dependable.
Retrieve Details When the Task Needs Them
Keeping every document in the active conversation is not the only way to make information available. A workflow can maintain an index of source names and locations, then retrieve the relevant sections as questions arise. Our RAG guide explains how retrieval can supply external material for an answer.
In the FAQ example, the agent can begin with the approved feature list and open the full compatibility document when it reaches a question about supported devices. It should preserve the restriction’s surrounding explanation, not just a matching sentence. The original source must remain accessible if the excerpt proves ambiguous.
Tool design affects what the agent sees. Anthropic’s tool guidance recommends meaningful, relevant responses and options for controlling output size. A useful tool response helps the agent understand the result while retaining identifiers needed for follow-up operations.
For our workflow, a document search result could include the title, revision, short excerpt and a stable reference. The agent can request more detail when needed. Returning every internal field from every matching record would add material without necessarily helping it answer the customer’s question.
Do not let concise output conceal incomplete retrieval. If a search returned only a subset of matches, label that limitation. A failed file read should not be represented as an empty document. The agent needs to distinguish evidence that says “not supported” from an absence of evidence caused by a broken tool.
Preserve enough provenance to check important statements. When the FAQ says a feature works offline, a reviewer should be able to trace that claim to the current source and inspect its qualifications. A citation to a broad homepage is less useful than the exact document section that establishes the behavior.
Use Summaries and Memory Without Losing the Evidence
Longer work often needs a compact account of progress. A useful working note records the task, decisions already made, open questions and source references. It should distinguish confirmed facts from assumptions, and completed actions from plans that have not been carried out.
Anthropic’s developer memory tool provides one implementation pattern: Claude requests operations on persistent memory files, and the application executes them against storage it controls. Information can be brought back into active context when needed instead of remaining in every request.
For the FAQ, a working note might say that the launch scope was checked against revision three, while the accessibility answer still needs confirmation. It should include the relevant source reference. Writing “accessibility confirmed” because an earlier assistant suggested it would convert an unverified statement into misleading project memory.
Compaction serves a different purpose. The compaction documentation describes replacing older conversation turns with a summary so a long-running task remains manageable. A compacted conversation supports continuity; it is not a substitute for retaining original evidence outside that summary.
Before summarizing our research session, preserve the approved scope, the exact unresolved conflict and the references needed to recheck both sides. Do not preserve every discarded phrasing attempt. After the handoff, ask the assistant to reopen a consequential source rather than assuming the summary reproduced every exception correctly.
Selective clearing is another option. Anthropic’s context-editing documentation describes removing older tool results under configured rules. Clearing raw results and summarizing a conversation are distinct operations. In either case, decide which evidence must remain retrievable before discarding it from the active working set.
A practical handoff test is to resume with only the working note and its referenced materials. Can the assistant identify what remains unresolved and continue without repeating completed research? If it cannot, improve the note’s structure rather than adding an unfiltered transcript back into every session.
Keep Evidence Separate From Authority
A retrieved document can help answer a question without being allowed to redefine the task. A customer email may contain a request, a product claim or an instruction addressed to automated systems. Those are different kinds of content, and the application should preserve that distinction.
For our example, a source saying “publish this announcement immediately” does not override the owner’s instruction to prepare a draft. Labeling the source as external can help communicate its role, but a label alone does not enforce permissions. The prompt injection guide covers how source material can try to redirect an agent.
Separate maintained project rules from the agent’s own working notes. Allow corrections through an appropriate owner-controlled process, while keeping observations and proposed rule changes clearly identified. An assistant should be able to suggest that the FAQ policy needs updating without quietly granting itself the authority to publish.
Context management also does not require giving the agent access to every account. The FAQ can begin with approved documents and a limited source connection. Add a tool or broader access when the task requires it, with the application enforcing the relevant boundary.
Test Whether Better Context Improves the Result
Use a small collection of tasks with answers you can verify. For the launch FAQ, prepare cases involving a superseded feature, an unresolved eligibility question and a current restriction. These are proposed tests for this illustrative workflow, not reported benchmark results.
Compare a baseline that supplies a large unorganized folder with a version that supplies a clear brief, labeled sources and targeted retrieval. Keep the model and intended task consistent when comparing the context strategies. Repeat important cases because one successful response does not establish repeatable performance.
Check factual accuracy, missing qualifications and source traceability. Also record unnecessary tool calls, time spent and the human effort needed to correct the output. A shorter context is not an improvement if it excludes the document that would have prevented a false claim.
Change the evidence deliberately. Replace the approved brief with a new revision, move the key restriction deeper into a document or make one source unavailable. The assistant should recognize the update, preserve the restriction or report the coverage gap. Our evaluation guide explains how outcome checks help assess these behaviors.
For longer tasks, include a summary or memory handoff in the test. Confirm that the resumed agent still distinguishes approved facts, historical information and unanswered questions. Context engineering works when the assistant has the right information at the right stage, and its answer remains connected to evidence that someone can inspect.
Kevin Weil’s Warp Sciences will build U.S. chemistry labs for drug developers, combining expert scientists with AI and progressively automated workflows.
