Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Repair Windows errors before they cause bigger problems3Scan for outdated or missing drivers - takes under a minuteiTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Preventing context bloat starts with identifying what is growing: tool definitions, intermediate tool results, or older conversation history. Tool search, code-mediated tool execution, context editing, and subagents address different pressures; a subagent can give work its own model-visible context, but does not automatically isolate application state, files, or authorization.
Diagnose what is filling the context window
Tool-heavy agents usually accumulate context in three places. Anthropic’s tool-context documentation identifies tool definitions and accumulated tool_result blocks as direct consumers of the context window. Long conversations can also retain results that were useful earlier but no longer matter.
- Definitions: Schemas and descriptions for many available tools occupy space even when a request needs only a few of them.
- Intermediate results: Tool outputs passed back through the model-visible conversation can accumulate across a chain of calls.
- Stale history: Earlier results may remain in the conversation after the agent has extracted the information it needs.
These causes are distinct. A smaller tool catalog does not remove a large result already in the conversation, and deleting old results does not make a large upfront catalog cheaper. Measure or inspect each source before choosing a control.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Match the control to the source of context pressure
| Pressure or need | Candidate control | What changes | Trade-off |
|---|---|---|---|
| Many tool definitions, few needed for a particular request | Tool search | Definitions can stay out of the initial prompt and be retrieved when needed. | Discovery adds a lookup step and depends on selecting the right tools. |
| Repeated chains of small tool calls | Programmatic tool execution | A script or execution environment can perform several calls without sending every intermediate result through the model. | It changes the execution pattern; not every workflow is safe or practical to convert. |
| Results that are no longer useful | Context editing | Old tool-result blocks can be removed after they have served their purpose. | The runtime needs a policy for deciding what is safe to discard. |
| Stable definitions repeated across requests | Prompt caching | Repeated-input costs may be reduced. | Caching does not reduce the number of tokens present in the context. |
| Independent work that benefits from parallelism or a separate model-visible conversation | Subagent | A delegated worker can focus on a bounded task and return a result to the coordinator. | Delegation adds orchestration and result-merging work; it does not guarantee savings for every task. |
Anthropic’s documentation suggests tool search as a rough starting point when a toolset grows beyond roughly 20 tools or baseline context becomes noticeable, context editing when conversations are long enough that prior results become irrelevant, and programmatic calling for repetitive small-call chains. These are vendor recommendations, not universal thresholds. It also recommends caching stable definitions in its high-volume starting point. See Manage tool context for the current guidance.
Use tool-scoped subagents for bounded, independent work
A subagent is most useful when its work can be stated as a clear question or deliverable, can proceed independently, and benefits from a narrower role or tool set. OpenAI’s multi-agent guide recommends clear independent tasks and expected results; short tasks or sequential steps that depend on the main agent’s immediate reasoning generally belong in the main agent instead.
Write a delegation contract
Before launching a worker, specify the boundaries in its task rather than relying on the label “subagent” to constrain it:
- Independent question: State the specific question or work product, and what the coordinator already knows that is relevant.
- Role: Give the worker a focused responsibility, such as inspecting a subsystem or checking a defined set of claims.
- Tool and data access: Name which tools, repositories, or data the worker may use. Exclude capabilities it does not need.
- Completion criteria: Define what counts as done, including any required checks or limits.
- Return format: Ask for a concise result with findings, relevant evidence, unresolved issues, and no unnecessary transcript.
Anthropic’s presentation on advanced subagent patterns likewise describes well-defined roles, tool-access levels, and success criteria. A narrow tool boundary can reduce accidental scope and make the worker’s job clearer; it is not itself proof that total runtime context, latency, or cost will fall.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Merge the result deliberately
The coordinator should incorporate only what the main task needs: conclusions, supporting details, and uncertainties. Returning a full tool transcript can simply move context growth from the worker to the coordinator. If several workers touch shared files, coordinate ownership and changes; separate model context does not mean separate filesystem state.
Separate model-visible context from application state
“The subagent has its own context” can mean that it has a separate model-visible conversation. It does not necessarily mean that it has an isolated application, execution environment, or security boundary. OpenAI’s Agents SDK context-management documentation distinguishes local context available to application code from context visible to the model. Local context can hold dependencies or state without being sent to the model.
The same documentation says derived wrappers in an SDK run share underlying application context, approval state, and usage tracking. Nested Agent.as_tool() runs do not automatically receive isolated copies of application state. OpenAI’s multi-agent documentation also notes that when a managed Agents API run has a configured environment, the coordinator and subagents share its filesystem. Treat each boundary separately:
- Model-visible conversation: What messages and tool results the model receives.
- Application state: Dependencies, state, approvals, and usage tracking available to the runtime.
- Execution resources: Filesystems and other resources workers may share through an environment.
- Authorization: Whether the implementation or server permits a requested operation.
Visibility filters can control which capabilities the SDK exposes, but they do not authorize model-generated arguments or resource selections. Enforce function-tool decisions in implementation code or with guardrails and approvals where appropriate. MCP servers must authorize their own protected operations. A narrower list of visible tools is useful, but it is not a substitute for enforcing access at the point where an operation executes.
Choose the runtime boundary you need
Runtime choice affects who owns orchestration, storage, tool execution, and the execution environment—not just how many agents you create. OpenAI describes the options in its Agents guide as follows:
- Agents API: A managed harness for long-running tasks.
- Agents SDK: An application-controlled agent loop with tools and handoffs.
- Responses API: A direct integration route.
Choose according to the control your application needs over those responsibilities, then verify the state and isolation behavior of the specific runtime and version you deploy. These product descriptions and behaviors can change; the cited documentation was accessed on October 7, 2026.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Interpret token-savings examples narrowly
Anthropic’s engineering article, “Code execution with MCP: building more efficient AI agents,” reports an example that goes from 150,000 tokens to 2,000 tokens and describes a 98.7% saving. The arithmetic is consistent with a 98.7% reduction in token count, but the article pairs that percentage with time and cost savings. It is a vendor-reported illustration of a particular pattern, not an independently established benchmark for arbitrary workloads or proof that subagents alone deliver the same result.
Measure your own representative tasks. Compare the context consumed by definitions, intermediate results, and retained history; also record lookup steps, execution time, and the amount of result synthesis your coordinator must do. The available sources do not establish universal token, latency, or monetary savings from delegation.
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteWindows 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 reinstallQuick Recap
A practical decision sequence
- Inspect the prompt and conversation: Determine whether definitions, tool outputs, or old history dominate context use.
- Keep unused definitions out: If a large catalog is the issue, test on-demand tool discovery and account for its additional lookup step.
- Move repetitive processing out of the conversation: For suitable chains, have code or an execution environment handle intermediate calls and return only the useful result.
- Discard obsolete results: Define when results can be removed and retain any facts needed for later reasoning in a deliberate, compact form.
- Delegate independent work: Set a role, tool boundary, completion criteria, and concise output contract; keep short or dependent work with the coordinator.
- Check actual isolation: Review model context, application state, files, approvals, usage tracking, and server authorization as separate concerns.
- Evaluate the whole workflow: Compare context, orchestration overhead, and measured performance on your own workload rather than assuming that an architectural label guarantees improvement.
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.
