Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errorsThe fastest way to cut token use is to measure the entire workflow, then remove unnecessary delegation, oversized tool context, repeated prefixes and obsolete history. Count input, output, reasoning, tool results and retries for the root agent and every subagent. Apply one change at a time, and compare total task cost and task quality—not merely a cache-hit percentage or a single model call.
What token usage includes in a multi-agent workflow
A multi-agent run spends tokens in more places than the final answer. The root agent’s instructions and conversation history are inputs; each specialist receives its own instructions and assigned context; tool definitions and tool results may be inserted into several turns; generated text, tool-call arguments and reasoning are outputs; and retries repeat some or all of that work.
As an Amazon Associate I earn from qualifying purchases.
OpenAI’s observability and usage guide identifies instructions, tool definitions, conversation history, user input, files or images and tool results as possible input contributors. Generated text, tool-call arguments and reasoning contribute to output usage; for the documented OpenAI API, reasoning tokens are billed as output tokens. Provider usage fields and traces are useful, but some fields can be best-effort or delayed, so treat an apparently complete trace as an accounting aid rather than an immutable invoice.
Your optimization target is the complete unit of work: all root and subagent calls, retries, tool or sandbox calls, and relevant third-party service charges. A workflow that makes one inexpensive model call but retries four times may cost more than a carefully designed two-agent run. Likewise, a cache-hit percentage can look excellent while the uncached requests, output and tool traffic still dominate the bill.
#1 Best Overall
- [Ultra-Lightweight Ear Buds Designed for Small Ears] Each earbud weighs only 3.7g and features a compact, ergonomic in-ear design made especially for small ears. Secure, low-profile, and comfortable for workouts, all-day wear or overnight sleeping.The ultra-slim profile sits flush in your ear with no hard edges poking or rubbing, so you can even sleep on your side without discomfort.
- [Immersive Stereo Sound with TOZO OrigX Technology] TOZO OrigX tuning delivers clear vocals, balanced mids, and natural stereo sound for music, podcasts, and videos.
- [Long Battery Life for Daily Use] Get up to 7 hours of playtime on a single charge, with up to 32 hours total using the charging case—ideal for workdays, commuting, and extended listening sessions.
- [Bluetooth 5.3 & Stable Connection] Bluetooth 5.3 provides fast pairing, stable wireless performance, and reduced dropouts as you move around home or office.
- [Deep Bass with Clear Vocals] High-performance drivers produce punchy bass while keeping vocals clean and detailed for everyday listening.
Step 1: Establish a baseline before changing the design
Capture a run-level record before optimizing. Give every task a correlation ID and attach that ID to root calls, subagent calls, tool invocations and retries. Store enough detail to answer both “where did tokens go?” and “did the result remain correct?”
- Call identity: model, provider, timestamp, root or subagent role, parent call and retry number.
- Input usage: uncached and cached input where the provider reports both, plus the major components (instructions, tools, history, files and tool results).
- Output usage: visible text, tool-call arguments and reasoning when exposed.
- Workflow events: delegation decision, tool calls, errors, retries, timeouts and compaction or context-editing events.
- Outcome: task success, evaluator score, human correction rate, latency and any external service cost.
OpenAI’s Agents guide and its usage documentation describe the categories and tracing approach. Use the equivalent usage fields and trace spans in another provider’s SDK. If a provider reports delayed usage, reconcile the final numbers after the run settles.
| Baseline question | Useful calculation | Why it matters |
|---|---|---|
| Which calls consume most input? | Input tokens by role and component | Shows whether history, schemas, files or duplicated instructions are the main problem. |
| Which calls consume most output? | Output and reasoning tokens by role | Identifies verbose specialists, excessive deliberation and retry loops. |
| What does one completed task cost? | All calls + retries + tool and external costs | Prevents optimizing a single call while total workflow cost rises. |
| Did quality change? | Pass rate, evaluator score and correction rate before versus after | Token savings are not useful if required evidence or decisions disappear. |
Run a representative sample rather than a single easy prompt. Record the distribution of task sizes and failure modes, because a change that helps short requests may hurt long, tool-heavy cases.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Step 2: Use only as many agents as the work needs
More agents do not automatically mean fewer tokens. Each subagent can repeat policy text, schemas and context, then produce a report that another agent must read and synthesize. Parallelism can shorten elapsed time or improve coverage, but it can also increase total usage.
OpenAI’s Multi-agent API guide recommends subagents for concrete, independent workstreams and warns that they may be less beneficial for a short task, a single ordered chain of reasoning, frequent writes to shared mutable state or work already dominated by one slow external operation. Its exact warning is: “Note that adding subagents can increase token usage, and may not be as beneficial for tasks that depend on a single ordered chain of reasoning, require frequent writes to shared mutable state, or are already dominated by one slow external operation.”
Keep one agent for sequential work
Use one agent when each step depends on the previous step’s evolving state: for example, inspect a file, decide on a change, apply it, run a test and revise the change. Splitting every step forces repeated handoffs and creates synthesis overhead. A single agent can retain the working state once instead of serializing it into multiple reports.
Split genuinely independent tracks
Delegation is a better fit when tracks have clear boundaries and can run without shared mutable state. Examples include examining separate codebase areas, checking independent source sets or generating parallel hypotheses that a final reviewer can compare. Give each specialist a bounded assignment, the smallest relevant context and only the tools it needs.
Set a delegation threshold
Make the root agent justify a subagent call with a simple rule: the work must be independent, large enough to offset setup and synthesis, and useful in parallel. For small requests, keep the work local. For a slow external operation, optimize that operation or wait once rather than multiplying agents around it.
Rank #2
- 【IP68 Waterproof Multifunctinal Smart Watch】Torjalph fitness activity tracker smart watch is IP68 waterproof technology ,you can wear it when swimming,handwashing,it's rain-proof which means that you perform multiple exercise modes outside without worrying about sweat or rain corroding your fitness smart watch,and this watch equipped with a 230 mAh large capacity battery magnetic charging ,and this sports activity tracker watch can be used 5-7 days If fully charged.
- 【Larger 1.83'' Full Screen Watch & Multi-dial Selection】1.83 inch full HD color touch screen smart watch for men / women provides excellent 240*284 resolution image quality ,This Bluetooth smart watch offers a better touch experience and remarkable immersion and ,more than 200 watch faces for you to choose from App"Da Fit ", you can change the dial style according to your mood to wear keeping the world on your wrist
- 【Accurate Activity Tracker Smart Watch 】sports watch activity tracker uses sweat and dust resistant breathable silicone strap, our Fitness Tracker Smart Watch for Android iOS Phones supports 12 sports modes,walking,running,bicycle,swimming,badminton,basketball,football,rope skipping. This bluetooth calls smart watch for men women is lightweight and stylish, bringing more convenient and relaxing sports experience while also providing more protection.
- 【Your 24 Hour Personal Daily Health Monitor】Fitness watch for women men full touch screen IP67 waterproof bluetooth smartwatch for Android iOS Phones with heart rate blood pressure monitor multifunction Andriod smart watch with steps counter sleep monitor sports fitness watch, monitoring blood oxygen saturation allows you to understand your physical condition in real time to track fitness progress ,to know your health status in time.
- 【More Functions Bluetooth Call Smart Watch】This fitness activity smart watch full screen for iPhone and Android phones includes a bunch of new features including:The bluetooth smart watch can hang up the phone, receive and display WhatsAPP /Text /Facebook /Twitter /Messenger reminder.Support sedentary reminder, alarm clock, stopwatch, timer,weather information,music control, remote camera control compatible with Android 5.0 and above or iOS 9.0 and above smartphones.
Return a compact handoff
Require specialists to return a fixed structure such as:
- Decision: the recommended action in one or two sentences.
- Evidence: only the observations that support the decision, with identifiers or links required for verification.
- Risks: unresolved uncertainty or conditions that would change the recommendation.
- Next action: the exact step the parent agent should perform.
This is a design recommendation, not a guaranteed percentage saving. Over-compressing a report can force the parent to call the specialist again, so keep the evidence needed to make the next decision.
Step 3: Reduce tool-definition and tool-result overhead
A large tool catalog can consume input tokens before the model has chosen a single tool. Tool results can be just as expensive: every verbose response inserted into history becomes context for later turns, and often for the parent agent as well.
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 →Load tools on demand
Expose a small core set initially and defer specialized schemas until the task requires them. OpenAI documents deferred tool loading in its prompt-caching guide. Anthropic describes a similar tool-search pattern in Manage tool context: definitions stay out of the context until a search identifies relevant tools. The lookup adds a turn and therefore some latency, but it can lower the baseline context for workflows that normally use only a fraction of a large catalog.
Return only fields needed next
Design tool responses as task-specific records rather than raw dumps. For a database query, return selected columns, row counts and a bounded sample instead of the entire table. For a file search, return matching paths, line ranges and short excerpts. Put large artifacts in durable storage and pass an identifier plus a retrieval instruction when the next step truly needs the full content.
Batch repetitive operations programmatically
When a chain makes many small, predictable calls, use a provider-supported programmatic tool-calling mechanism. Anthropic’s tool-context guide describes executing multiple calls in one script so intermediate results do not become separate conversation-history round trips. This can reduce repeated model turns, but it requires careful script validation, permissions and error handling. Keep a limit on batch size and return a concise aggregate with per-operation status.
Delete obsolete tool results
Once a result has been converted into a durable fact—such as “test suite passed at commit identifier X”—remove the raw output if later reasoning does not need it. Context editing features differ by provider; implement an explicit retention policy even when the API offers automatic editing. Preserve identifiers, errors, assumptions and evidence locations so an agent can retrieve details instead of guessing.
Step 4: Make prompt caching work for repeated prefixes
Prompt caching targets the cost of processing repeated input; it does not shorten the context. Anthropic states this distinction directly: “Prompt caching doesn’t reduce the number of tokens in context, but it reduces what you pay for them on subsequent requests.” Cached content still occupies the model’s context window and still contributes to input usage.
Rank #3
- [Low Noise High Fidelity] New bee guitar cord with 22AWG oxygen-free-copper center conductors provides enhanced signal transfer, the PVC insulation material and the high-density stranded copper double shield provide noise rejection for quiet operation
- [Ideal for Choice] New bee 1/4 straight and right angle metal connectors, ensures a secure connection in any instrument, pedal, or amp jack. L shape great fro on stage guitar stands
- [Wide Compatible] Suitable for mixer, power amplifier, microphone, effector, decoder, equalizer, electoronic organ, electric guitar, electric bass and other electronic equipments. Widely used in the installation of smart home, karaoke speaker, professional stage audio, etc
- [Easy Disassemble] Making incidental repairs quick and easy! Very nice, flexible cable
- [Dont Hesite to Order] 24 hours customer services and professional technology team are standing by
Keep the rendered prefix identical
OpenAI’s prompt-caching documentation explains that reuse depends on a matching rendered prefix, not merely similar wording. Put stable policy instructions, tool definitions and reusable reference material first; append the changing user request, task identifiers and recent results afterward. A change to content or relevant settings before the cache breakpoint can prevent reuse.
Do not generate timestamps, random IDs or changing diagnostics inside the stable prefix. Normalize serialization, tool ordering and whitespace where the provider’s interface makes those part of the rendered request. Keep separate prompt versions for materially different policies instead of adding conditional clutter to every request.
Account for cache eligibility and timing
Minimum cacheable lengths, retention periods, breakpoints and prices vary by model and API version. Confirm the current rules before forecasting savings. Google says implicit caching is enabled by default for Gemini 2.5 and newer models and recommends placing large common content early and sending similar prefixes close together; its current guidance is at Gemini context caching (page listed as updated 2026-10-09 UTC). These defaults are version-specific and can change.
Measure dollars and tokens together
Compare uncached input cost, cached input cost, output and reasoning, number of calls and total task cost. A high cached-input share can coexist with more total tokens if the workflow added agents or retries. After changing a prefix, check both cache reuse and quality; a cache-friendly prompt that omits a necessary constraint is not an optimization.
Step 5: Bound long-running context
Long conversations accumulate stale tool results, superseded plans and duplicate explanations. Before adding another specialist, inspect whether the root agent is carrying history that no longer affects the next decision.
Use explicit context editing
At safe checkpoints, remove raw tool outputs and old intermediate drafts while retaining current requirements, decisions, identifiers, unresolved questions and evidence locations. Anthropic documents context-management approaches, including removing old tool results, in Manage tool context. A retrieval pointer is safer than silently discarding information: the next agent can fetch the original artifact if a disputed detail matters.
Use server-side compaction when supported
OpenAI’s Compaction guide describes threshold-triggered server-side compaction for the Responses API. The service returns a compaction item that carries forward key prior state in fewer tokens. Configure a threshold appropriate to the model’s context window, then test the resulting state on representative tasks.
Recommended Free Tools
Compaction can change the rendered prefix and reduce cache reuse on the first request afterward. Compare total input cost before and after compaction rather than assuming that a shorter context always produces an immediate bill reduction. Validate that the compacted state preserves constraints, decisions, identifiers and unresolved work; the mechanism itself is not a universal quality guarantee.
Rank #4
- Digital Family Calendar: BSIMB digital calendar wall touch screen puts an end to your chaotic life. Designed for busy families, the digital calendar keeps you organised and in step with your family. Assign different colours and tasks to each family member to make family scheduling easier and more intuitive.
- Easy Setup and Auto-Sync: Plug in, connect to WLAN and auto-syn your electric calendar in minutes. Digital wall calendar compatible with Google calendar, iCloud calendar, Outlook calendar, Cozi calendar and Yahoo calendar. Add events directly on the device or through the ecalendar app. Electronic calendar stays current whether you're at work or traveling—no manual updates needed.
- All-In-One Home Assistant: Smart calendar builds habits and encourage independence with personalised chores. Display daily meal plan to reduce the daily ‘What's for dinner?’ stress. To-do lists keep track of everything you want. Turning chores into valuable achievements with stars and rewards, making life easier and more enjoyable.
- Manage Anytime Anywhere: BSIMB electronic calendar for family stay connected with the powerful ecalendar app. Access and manage your digital wall calendar anytime, anywhere with free features like stars & rewards, photo screensaver, chore charts, meals plans, to-do lists etc. Electronic planner paid features like Magic Import tailor advanced tools for users with special needs.
- Stay in Touch with Loved Ones: Share photos or videos from your phone to electronic wall calendar with app, instantly update your life moment, and keep in touch with your family, especially those who are far away; you can also surprise your smart calendar digital wall by sending him a wish card; it's also great for keeping tech-illiterate elders in touch with the youngsters.
Compact by meaning, not by character count
A useful summary records what must remain true, what has been decided, what remains open and where supporting evidence lives. Do not summarize away a security constraint, an API parameter, a failing test or a user’s exception. If a later step needs exact text, retain a reference to the source rather than an approximate paraphrase.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Step 6: Standardize prompts and handoffs
Repeated task types often grow separate prompts that drift apart and duplicate policy text. OpenAI’s practical guide to building agents describes a flexible prompt template with policy variables instead of many independent prompts. One maintained template is easier to evaluate and can keep the stable prefix identical across requests.
Separate policy, role and task data
Place invariant safety and formatting rules in a versioned policy block. Define the specialist’s role and output schema in a stable role block. Supply user data, retrieved documents and run-specific constraints afterward. This layout improves cache reuse and makes it clear which changes should trigger a new prompt version.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Set output limits that match the decision
Ask for a bounded number of findings, a maximum excerpt length and an explicit “no finding” value. Require citations or artifact identifiers for claims that need verification. Avoid asking every specialist to restate the assignment, all source text and the complete reasoning chain; request the evidence the parent needs to act.
Prevent duplicate synthesis
Choose one component to synthesize. If each specialist writes a polished essay and the root agent rewrites all of them, output tokens multiply without adding evidence. Specialists should report findings; the root agent should make the final decision and communicate it once.
Choosing an optimization by bottleneck
| Approach | Targets | Trade-off | Best fit |
|---|---|---|---|
| Fewer or more selective agents | Duplicate instructions, calls and synthesis | Less parallel coverage or more elapsed time | Small, sequential or externally constrained tasks |
| On-demand tool discovery | Schemas loaded before they are needed | Extra lookup turn and latency | Large catalogs where only a subset is usually used |
| Programmatic tool calling | Repeated tool-call/result round trips | Requires supported execution and robust scripts | Repetitive chains of small operations |
| Prompt caching | Reprocessing repeated identical prefixes | Does not shrink context; eligibility and pricing vary | Stable instructions, schemas and reusable references |
| Context editing or compaction | Stale history and oversized prompts | May remove useful context or disrupt cache reuse | Long-running conversations with obsolete results |
| Templates and structured handoffs | Prompt duplication and verbose reports | Over-compression can omit constraints or evidence | Repeated tasks and bounded specialist assignments |
These are workflow choices, not guaranteed savings percentages. The winning intervention is the one that reduces the measured dominant component while preserving the required outcome.
A practical implementation sequence
- Instrument one representative workflow. Record root, subagent, tool and retry spans; capture input, output, reasoning and outcome metrics.
- Rank contributors. Separate duplicated instructions, tool schemas, tool results, history, generated reports, reasoning and retries.
- Remove avoidable work. Disable unnecessary retries, stop passing stale results and collapse agents that perform a short sequential chain.
- Right-size each specialist. Give it only bounded context, required tools and a structured response contract.
- Fix tool traffic. Defer rarely used schemas, batch repetitive calls where supported and return only fields needed by the next step.
- Stabilize repeated prefixes. Move changing data after invariant instructions and schemas; verify actual cache reuse under the provider’s current rules.
- Bound history. Edit old tool results or enable threshold-triggered compaction, preserving constraints and evidence pointers.
- Run quality checks. Use the same evaluation set, compare correctness and correction rate, and inspect failures caused by missing context.
- Adopt the change only if the whole task improves. Include latency, external costs, retries and cache effects—not just token totals for one call.
Common failure modes and fixes
“The cache-hit rate is high, so costs must be lower.”
A cache-hit share describes eligible repeated input, not total workflow economics. Check added subagent calls, output, reasoning, retries and tool costs. Compare the complete task before and after.
Free tools Windows power users keep installed
One-click scans. No signup required.
“Every task gets a specialist for maximum quality.”
Specialists add setup, duplicated context and synthesis. Route short or tightly ordered work to one agent; reserve parallel tracks for independent, bounded work.
“Send the entire tool response so the model has all the facts.”
Raw responses enlarge every later prompt. Return a bounded, typed result and retain a retrievable artifact for details that may be needed later.
“Summarize aggressively at every turn.”
Repeated summaries can consume more tokens than the original state and may erase exact constraints. Summarize at meaningful checkpoints, preserve identifiers and test summaries against the original task.
“Compaction is free context reduction.”
Compaction can alter cache reuse and remove information. Measure post-compaction input cost and run quality tests that exercise unresolved work, exceptions and tool failures.
“A shorter report is always better.”
If the parent lacks evidence, it may call the specialist again or make an unsupported decision. Set a compact schema that includes the decision, necessary evidence, risks and next action.
How to verify that the optimization worked
Use an before-and-after report for the same workload mix. For each task, compare total input (cached and uncached), total output and reasoning, number of model calls, tool calls, retries, latency, external charges and quality score. Break the result down by task size and failure type; an average can hide a severe regression in long-running cases.
Inspect traces for accidental regressions: a supposedly deferred tool catalog loaded on the first turn, a cache prefix changed by a timestamp, a compaction event followed by a full-history resend, or a specialist retrying because its output schema omitted an identifier. Keep the old configuration available for rollback until the new design has passed the same evaluation set over enough production traffic to expose variance.
Provider behavior changes. OpenAI documents agent integration at Agents, multi-agent trade-offs at Multi-agent, caching at Prompt caching, compaction at Compaction and accounting at Observability and usage. Anthropic’s approaches are documented at Manage tool context; Google’s current caching behavior is documented at Context caching. Check the applicable model and API version before relying on thresholds, defaults or pricing.
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 minutePC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Quick 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.
