Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Context engineering is the work of deciding what an AI agent can use at each step—not just writing its prompt. It means selecting and maintaining the instructions, tool definitions, conversation history, retrieved evidence, and saved state that enter the model’s active context. The goal is to give the agent enough relevant information to act well without filling its limited working space with material that does not help its next decision.
What is context engineering?
Context engineering is the ongoing design of the information supplied to a model during inference. Anthropic defines it as strategies for curating and maintaining the optimal set of tokens, including information that reaches the model outside the prompt itself. Anthropic’s guide to context engineering frames the task as choosing the context configuration most likely to produce the desired behavior.
![Claude AI for Beginners Bible: [5 in 1] The Ultimate Guide to Automate Your Work, Save Hours Every Week, and Use AI for Real-World Results](https://m.media-amazon.com/images/I/415+fSJacsL._SL160_.jpg)
As an Amazon Associate I earn from qualifying purchases.
Prompt engineering focuses on instructions and wording. Context engineering includes that work, but also asks what else the agent needs now, what it can fetch later, and what should be summarized or discarded. In a multi-step agent, the answer changes as tool calls return results and the task progresses.
What counts toward an agent’s context?
The context window is the text a model can reference while generating a response, including the response itself, according to Anthropic’s context-window documentation. The active context may contain:
#1 Best Overall
- System and developer instructions, plus the current user request and constraints.
- Tool definitions that tell the model which actions are available and how to call them.
- Earlier messages, decisions, and assistant-generated output.
- Tool results, such as file contents, search results, or API responses.
- Retrieved documents or other external evidence relevant to the current step.
- Output tokens generated by the model; depending on provider and model, other configured or reasoning tokens may also count toward the limit.
Tool definitions and returned results consume context too. As a result, an agent can run into context pressure even when its visible conversation is short. Exact limits and counting rules vary by provider and model, and change over time; check the target model’s current documentation rather than assuming a capacity applies universally.
Why isn’t a larger context window enough?
A bigger window can accommodate more material, but it does not make every included token useful. Anthropic describes context as a finite resource with diminishing returns, and its documentation warns that accuracy and recall can decline as token counts grow. Irrelevant content competes with task-critical instructions and evidence for the model’s attention.
Position can matter as well as volume: information buried in a long history may be less useful to the next decision than a concise, current statement of the goal and constraints. Context design therefore means managing relevance and organization—not simply maximizing the number of tokens supplied.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problemsHow do I manage context for a long-running agent?
Start by identifying what the agent needs on its next inference step. Keep stable instructions and current constraints available; bring in past decisions, evidence, or detailed tool output when they affect the next action. For long tasks, plan explicitly for state recovery: the agent may need to continue after older messages have been compressed or a session has ended.
Retrieve evidence when it is needed
Use retrieval to bring relevant external material into the active context at the point it can inform a decision. Embedding-based retrieval is one common pre-inference approach. For capable agents, just-in-time strategies can fetch information as the task calls for it, rather than loading an entire corpus at the start.
Retrieval is most useful when the source material is too large or changeable to keep in every request. The application still needs to retrieve the right passages and give the model enough context to interpret them; retrieval does not guarantee relevance or correctness by itself.
Compact history when the conversation is the problem
Compaction summarizes a long conversation so the agent can continue with a shorter history. A useful summary preserves the goal, constraints, decisions already made, unresolved questions, and implementation details needed to proceed. It should not flatten away distinctions that could change the next action.
Clear old tool results when outputs dominate
Tool-result clearing removes old, re-fetchable outputs from active context while retaining the fact that the call occurred. This can help when raw file reads, search results, or API responses have accumulated. Keep a result if it contains information that cannot be recovered or would be costly or risky to fetch again.
Use persistent memory for knowledge that must survive
Memory stores selected information outside the active context so it can persist across context resets or sessions. It is appropriate for durable facts, preferences, or project state that future steps genuinely need—not as a dumping ground for every interaction. In Anthropic’s described memory-tool approach, the application developer controls the storage backend.
Leave structured notes for long-horizon work
For a project that may span many steps or sessions, structured notes can record progress and make state easier to reconstruct after a reset. Focused subagents can also handle bounded subtasks with narrower context. Both approaches depend on good task decomposition and a clear handoff: the main agent needs to know what was done, what remains, and which findings matter.
Which context technique should I try first?
| Problem | Technique | What it changes |
|---|---|---|
| Too much external material to load at once | Selective retrieval | Fetches relevant evidence into context when needed. |
| Conversation history is growing too long | Compaction | Replaces older history with a summary of task-critical state. |
| Large, reproducible tool outputs are accumulating | Tool-result clearing | Removes old outputs that can be fetched again while preserving that the calls occurred. |
| Knowledge must persist across context resets or sessions | Persistent memory | Stores selected information outside the active context for later use. |
| A long task needs focused parallel work or recovery notes | Structured notes or focused subagents | Records progress or delegates a bounded task with a deliberate state handoff. |
These methods address different sources of context pressure and can be combined. Choose based on what is growing or being lost: history, tool output, external evidence, or state across sessions.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallHow should I evaluate a context strategy?
Test changes against the agent’s real workload and tool-use pattern. A technique that helps a web-search agent may not help one that repeatedly edits files or queries a database. Compare configurations under the same tasks and look for trade-offs in task performance, token use, latency, and reliability.
Best Value
- Identify where context growth comes from: message history, tool definitions, tool results, or retrieved material.
- Specify what each approach retains, removes, or stores outside the active context.
- Check whether the task needs persistence across turns, resets, or sessions.
- Measure whether the agent still uses the right evidence and preserves important decisions after summarization or clearing.
- Track engineering effort and control over any external storage or retrieval system.
Anthropic’s agent patterns cookbook recommends matching context-management primitives to the workload’s actual problem and testing clearing configurations against its tool-use pattern.
What do Anthropic’s evaluation results show?
In a 2025 announcement, Anthropic reported results from its own internal agentic-search evaluations: combining its memory tool with context editing improved performance 39% over its baseline, while context editing alone improved it 29%. In a separate 100-turn web-search evaluation, Anthropic reported an 84% reduction in token consumption using context editing. These are vendor-reported results for Anthropic’s evaluations, not universal guarantees or independent replications; they do not establish that the same technique or gains will transfer to another agent or workload. Anthropic’s context-management announcement
A practical starting decision
- If irrelevant documents are entering every request, begin with selective retrieval.
- If old conversation turns are the main source of growth, test compaction and validate that its summaries preserve decisions and open issues.
- If repeatable tool responses dominate, test clearing those results and retain anything that cannot safely be fetched again.
- If work must continue across sessions, define what belongs in persistent memory or structured progress notes.
- If a subtask can be isolated, delegate it only when the task boundary and return handoff are clear.
Make one targeted change at a time, then compare it on representative tasks. The best context is not the largest one: it is the smallest, clearest working set that gives the agent what it needs for the next decision.
Recommended Free Tools
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
