Source: https://internal-agents.com/lessons/load-tools

[← Lessons](https://internal-agents.com/lessons)

05 / Tools and context

# Two ways to discover tools

Cloudflare searches tool schemas; Sentry connects to providers on demand.

11 September 2026 · Updated 16 September 2026 · 2 min read

## Two discovery paths

Cloudflare measured roughly 15,000 tokens for 34 GitLab tool schemas before the model started a task. Its portal now exposes search and execution functions in place of every upstream definition. The agent searches for a capability, then uses the selected tool. [[1]](https://internal-agents.com/lessons/load-tools#source-1)

Sentry's Junior initially connects to no Model Context Protocol (MCP) provider. MCP gives the agent access to tools. Junior requests a lookup for a named provider, connects to it, and discovers its tools when needed. [[2]](https://internal-agents.com/lessons/load-tools#source-2)

| Design | What is available initially | What discovery adds |
| --- | --- | --- |
| Cloudflare portal | Search and execution functions | The tool schema needed for a particular operation |
| Sentry Junior | A way to request a provider's tools | A provider connection and its available tools |

The two designs defer different work. Cloudflare searches across a catalog of schemas; Junior first needs to reach the relevant provider. This puts connection failures and search failures at different points in the task.

## Discovery can fail before execution

Our interpretation is that these designs exchange upfront context for a dependency during the task. A search can return the wrong capability. A provider can be unavailable or fail authorization. A useful tool can also have a description that prevents the agent from finding it.

Those failures need to be distinguished from a tool that was found but returned an error. Cloudflare's context measurement supports the overhead concern; it does not establish an accuracy improvement. For a small, stable set of tools, discovery would add another operation whose benefit still needs measurement.

## Sources

1. [Cloudflare: Our internal AI engineering stack](https://blog.cloudflare.com/internal-ai-engineering-stack/) Tool schema overhead and portal-level search.
2. [Sentry: Building an intern](https://cra.mr/building-an-intern/) Progressive discovery and provider connections.

### Related items

[Cloudflare in the catalog](https://internal-agents.com/agents/cloudflare-ai-stack) [Sentry in the catalog](https://internal-agents.com/agents/sentry-junior)

[All lessons](https://internal-agents.com/lessons) [Next: Some steps do not need a model →](https://internal-agents.com/lessons/steps-without-a-model)