# Capability Bloat Evaluation

> Capability bloat evaluation is the audit of an agent that already has its tools and capabilities configured: find the ones that overlap or carry more privilege than the task needs, then remove, merge or scope them and re-run the tasks the agent must still cover. The count of tool names is not the measure; the effective privileges and the overlap in what each tool can do are.

**Domain:** CCARP-D3 · Integration (19% of the CCA-P exam)
**Canonical:** https://claudearchitectcertification.com/concepts/capability-bloat-evaluation
**Last reviewed:** 2026-10-16

## Quick stats

- **Audit actions:** Remove, merge, scope
- **Measure:** Effective privilege
- **Exam objective:** CCARP-D3-O1
- **Exam domain:** CCARP-D3

## The decision

Capability bloat is surplus tool surface: overlapping tools that blur the model's choice, and broad tools whose effective privileges exceed the task. Audit what each tool can actually do, then remove, merge or scope it.

**Learner decision:** Which capabilities should I remove, merge or scope, and how do I prove the agent can still do its job?

## What it is

A configured agent tends to collect capabilities: a second search tool added for one project, a broad action tool kept because it was convenient, a connector nobody removed. Capability bloat is the state where the configuration offers more overlapping or more privileged capability than the agent's tasks need. The cost is practical: the model has more near-duplicate tools to choose between, the tool definitions take up context, and each over-broad tool widens what a wrong choice can do.

Evaluating for bloat is different from choosing a tool surface in the first place. Choosing among built-in tools, custom tools, Skills and MCP is a design decision made before the configuration exists. This page covers the audit of a configuration that already exists: which capabilities overlap, which carry excess scope, and what to remove, merge or restrict without losing the tasks the agent must still complete.

## How it works

Start from the tasks, not the tool list. List the tasks the agent must be able to complete. For each tool, record what it can actually do (read, write, delete, send) and on which resources. Two tools with different names that can do the same thing are an overlap; one tool that can do far more than any task requires is excess scope.

Choose one action per finding. Remove a capability no task needs. Merge two overlapping capabilities into one with a clear description. Scope a broad capability down to the resources and verbs the tasks need. Keep a capability only when a named task depends on it.

Validate before and after. Produce a before and after capability matrix, then re-run the representative tasks against the reduced configuration. A removal that breaks a required task is a defect in the audit, not a success of the cleanup.

## Diagram

*Audit sketch: From capability list to validated configuration*

1. **List the required tasks.** The tasks the agent must still complete after the audit.
2. **Map each capability to its effective privileges.** What each tool can read, write or send, and on which resources.
3. **Mark overlap and excess scope.** Different names with the same reach, or reach beyond what any task needs.
4. **Remove, merge or scope.** One action per finding, each tied to a task or its absence.
5. **Re-run the tasks.** Confirm every required task is still covered by the reduced configuration.

## Exam trap

Judging bloat by the number of tool names, or keeping an overlapping capability for flexibility. The measure is effective privilege and overlap, and every removal has to be checked against the tasks the agent must still complete.

## Worked example: auditing a support agent's tool set

*Authored teaching example, not a bank question.*

A support agent has three search tools, a database tool that can write, an unrestricted HTTP tool, an email tool, a refund tool, an account-deletion tool and an address-update tool. Its stated tasks are answering order questions, drafting replies for staff to send, and updating a shipping address when a customer asks.

1. **Inventory by effect.** List each tool with what it can read, change and reach, not just its name. The database tool can write to any table, and the HTTP tool can contact any host.
2. **Find overlap.** Three search tools return similar results from different stores, so the model must guess which to call. Merge them into one search tool with a source parameter.
3. **Scope broad tools.** Give the database tool read-only access to the tables the tasks need. Remove the HTTP tool, or restrict it to a short allowlist of hosts. Remove the email, refund and account-deletion tools, because no stated task needs them: staff send the drafts, and refunds and deletions are handled elsewhere. Keep the address-update tool, which a stated task needs, behind an approval gate.
4. **Validate coverage.** Run the representative task suite before and after. Every task that passed before must still pass, and record any new tool-choice errors.

| Tool | Problem | Action | After |
| --- | --- | --- | --- |
| Three search tools | Overlap blurs the choice | Merge | One search with a source parameter |
| Database tool | Can write anywhere | Scope | Read-only on needed tables |
| HTTP tool | Reaches any host | Remove or restrict | Allowlist or removed |
| Email, refund and delete tools | No stated task needs them | Remove | Gone from the configuration |
| Address-update tool | Needed for a stated task, consequential | Keep and gate | Approval before execution |

The agent has fewer, clearer tools with narrower reach. Check afterwards that no retained tool lacks a task that needs it and that the effective privileges, not just the tool count, went down.

## Failure modes

### Counting tool names instead of effective privileges

- **Looks right:** The agent has only five tools, so the surface is small.
- **Why it fails:** One tool that can run any query or reach any host is wider than several narrow ones. Tool count hides reach.
- **Instead:** List what each tool can read, change and contact, and cut or scope by effect.

### Removing capabilities without tests

- **Looks right:** A tool was rarely used, so deleting it is safe.
- **Why it fails:** A rare tool may be the only path for an important task, and the loss shows up in production.
- **Instead:** Run the task suite before and after, and remove only what no task needs.

### Keeping overlap for flexibility

- **Looks right:** Several similar tools give the model options.
- **Why it fails:** Ambiguity among tools leads to wrong or inconsistent choices and spends context on definitions.
- **Instead:** Merge overlapping tools into one with clear parameters, or remove the weaker duplicate.

## How each certification frames it

### CCA-P: How CCA-P tests capability bloat

Objectives: `CCARP-D3-O1`

The Professional guide lists evaluating tool and agent configuration for capability bloat (PDF page 5, Domain 3). This is an audit of a configuration that already exists, which differs from choosing among built-in tools, custom tools, Skills and MCP when designing one.

Anthropic's context engineering guidance names bloated tool sets that cover too much functionality as a common failure and recommends a minimal viable set of tools. Its secure deployment guidance adds the privilege side: restrict filesystem, network and credentials to what the task needs.

## Comparison

| Finding | Typical action | Check afterwards |
| --- | --- | --- |
| No task depends on the capability | Remove it | Required tasks still pass |
| Two capabilities reach the same resources | Merge into one with a clear description | The model picks the merged tool for those tasks |
| One capability can do more than any task needs | Scope its verbs and resources down | Out-of-scope actions are refused, in-scope ones succeed |
| A capability is kept for flexibility only | Treat as removable until a task names it | No required task regresses |

## Decision tree

1. **Does a named task depend on this capability?**
   - **Yes:** Keep it, then check whether its scope is wider than that task needs.
   - **No:** Remove it.

2. **Does another capability already reach the same resources for the same task?**
   - **Yes:** Merge the two into one and describe it clearly.
   - **No:** Keep them separate.

3. **Can the capability do more than the task requires?**
   - **Yes:** Scope it down to the verbs and resources the task needs.
   - **No:** Leave its scope as it is.

## FAQ

### Q1. How is capability bloat different from choosing a tool surface?

Tool surface selection decides among built-in tools, custom tools, Skills and MCP when you design an agent. Capability bloat evaluation audits a configuration that already exists for redundancy and excessive scope.

### Q2. How many tools are too many?

Anthropic gives no fixed number. The guidance is a minimal viable set: every tool earns its place by a task that needs it, with no ambiguity about which tool applies. Measure tool-choice errors rather than counting.

## Sources and next steps

- [Claude Certified Architect - Professional Exam Guide v1.0 (July 2026), PDF page 5](https://claudearchitectcertification.com/exam-guide/ccar-p), section: Domains 3, 4 and 5, detailed objectives. Checked 2026-10-16.
- [Claude Certified Architect - Professional Exam Guide v1.0 (July 2026), PDF page 6](https://claudearchitectcertification.com/exam-guide/ccar-p), section: Domain 7: Developer Productivity & Operational Enablement, detailed objectives. Checked 2026-10-16.
- [Effective context engineering for AI agents (Anthropic engineering)](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents), section: Context as a finite resource; Compaction; Structured note-taking; Sub-agent architectures; Tool-set bloat. Checked 2026-10-16.
- [Securely deploying AI agents (Claude Code docs)](https://code.claude.com/docs/en/agent-sdk/secure-deployment), section: Threat model; Built-in security features; Security principles. Checked 2026-10-16.

- [CCA-P exam guide: Domain 3, Integration](https://claudearchitectcertification.com/exam-guide/ccar-p)
- [Agentic tool surface selection](https://claudearchitectcertification.com/concepts/agentic-tool-surface-selection)
- [Authentication and access control for agents](https://claudearchitectcertification.com/concepts/auth-and-access-control-for-agents)
- [Agentic tool design scenario](https://claudearchitectcertification.com/scenarios/agentic-tool-design)

---

**Source:** https://claudearchitectcertification.com/concepts/capability-bloat-evaluation
**Last reviewed:** 2026-10-16

**Evidence tiers**, 🟢 official Anthropic doc / API contract · 🟡 partial doc / inferred · 🟠 community-derived · 🔴 disputed.