Tool configuration: choosing the right context intervention
Tool configuration: choosing the right context intervention
This topic teaches CCARP-U3.T1.LO1: Evaluate tool/agent configuration for capability bloat. The decision is whether exposed capabilities and retained results help the task enough to justify their context and operating costs.
Diagnose before changing the catalog
A large tool count is a clue, not a diagnosis. Inspect what occupies the initial request, what accumulates during the conversation and what is repeatedly billed. Keep these observations separate.
| Observation | Candidate mechanism | What it does not establish |
|---|---|---|
| Mostly unused definitions dominate startup | Tool search | That every needed tool will be discovered within the latency target |
| A stable prefix is repeatedly charged | Prompt caching | That occupied context shrinks |
| Obsolete tool results accumulate | Context editing | That all old evidence is safe to discard |
The official documentation distinguishes these mechanisms and explains that they can compose. Tool search defers definitions until requested and adds a lookup turn. Caching reduces repeat-input cost without reducing context occupancy. Context editing removes old tool-result blocks once they have served their purpose.
Worked decision — synthetic support assistant
Requirements: an assistant has 80 available tools, usually needs two per task and must preserve the existing task-success rate while meeting its agreed latency target. These figures are exercise inputs, not vendor limits.
Evidence: traces show that unused definitions dominate the first request. This happens before result history exists.
Decision: pilot deferred discovery. It targets the measured startup load. Compare it with the current design on the same representative tasks and include the lookup in end-to-end timing.
Alternative ruled out for this cause: caching the complete catalog may improve repeat-input economics, but it does not remove the definitions from context. History cleanup addresses a later problem and cannot by itself solve initial definition bloat.
Failure and response: if task success drops because a needed capability is missed, inspect the discovery trace and repair that path before expanding the rollout. A smaller input does not compensate for a failed acceptance criterion.
Review record
- Name the measured source of pressure and retain the trace that demonstrates it.
- Explain how the selected mechanism changes that source.
- Record the required capabilities and the evidence that must remain available.
- Compare task success and end-to-end latency before and after the change.
- State an explicit rollback condition and owner.
These review steps are house implementation guidance. The vendor documentation establishes the mechanism; your workload establishes whether it is a good choice.
Source
Define tools — Claude documentation supports clear descriptions and consolidation of related operations.
Manage tool context — Claude documentation. Retrieved source text, hash and exact supporting excerpts are retained with this topic's claim ledger.