Adding a tool with Claude’s inline runner may preserve more of a conversation’s request body than changing the request’s tools[] array—but the striking 98.7% figure comes from one author-reported test, not an Anthropic guarantee. Chew Loong Nian’s September 25, 2026, Towards AI article reports up to 98.7% request-body reuse for a next request in a 40-turn conversation after using addTools(), versus no reuse in that test after editing tools[]. That result can inform implementation choices, but it does not establish universal behavior, lower latency, or lower cost.
What the 98.7% figure measures
The percentage refers to request-body reuse in the author’s reported 40-turn test: when a tool was added with addTools(), the next request reportedly reused up to 98.7% of its body. In the same comparison, editing tools[] reportedly produced no reuse for that next request. The article does not establish that this outcome will recur across runs, models, SDK versions, or workloads.
![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)
Request-body reuse is not itself a measurement of response latency, cache hit rate in every sense, token billing, or money saved. The cited article reports no separate result for those outcomes, so the percentage should not be translated into a cost or speed claim.
How the two approaches differ in the reported account
| Approach | How tools change | Reported effect on the next request | Evidence scope |
|---|---|---|---|
addTools() |
The article describes adding a tool through the inline tool runner during an active run. | Up to 98.7% request-body reuse. | Chew Loong Nian’s reported result in a 40-turn conversation; not independently reproduced in the available sources. |
Edit tools[] |
Change the tool list in the request. | No request-body reuse. | The same author-reported comparison; not proof that all direct tool-list changes always eliminate reuse. |
To explain the difference, the article says a modified tools[] array changes the cached prompt prefix, whereas addTools() appends a tool addition while preserving the earlier prefix. This is the article’s explanation of the reported behavior; the official Anthropic documentation located for this topic does not independently confirm the implementation details.
#1 Best Overall
What the article says about the inline runner
The article says Anthropic SDK 0.128.0 includes an inline tool runner with addTools() and removeTools() for changing available tools during a run. It also says the runner requires the beta flag inline-tools-2026-09-15 to be included explicitly because the runner does not add it automatically.
These details are version-sensitive. The official Anthropic prompt-engineering page provides general tool-use context but does not establish these methods, that beta flag, or the reported reuse result. Check current official SDK documentation and support information for your specific SDK, model, and platform before relying on the feature: Anthropic’s prompt-engineering documentation.
Rank #2
Choosing an approach for a long-running agent
If the available tools need to change mid-run, the reported result makes addTools() worth evaluating where the inline runner is supported. Do not choose it solely on the basis of the headline percentage: first verify that the method and beta flag are currently supported in your environment, then measure the behavior that matters for your workload.
- Confirm the current SDK version, model and platform support, and beta-header or parameter requirements in official Anthropic materials.
- Compare the approaches using the same conversation history, tool definitions, and request sequence; measure request-body reuse directly rather than assuming the reported result generalizes.
- Measure latency and billing separately if those are your reasons for changing implementation. The cited comparison reports request-body reuse, not those effects.
- Use
removeTools()only if your design needs to withdraw tools during an active run; its presence in the article’s feature description does not establish a reuse result for removal.
Chew Loong Nian’s article is the source for the feature description, beta-flag guidance, mechanism explanation, and single reported benchmark: Towards AI, September 25, 2026. The available evidence does not include an independent reproduction or an official Anthropic confirmation of the specific benchmark.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
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.
