Blog / Agent-ready tools

Agent-ready tools glossary: tool search, deferred tools, code mode, peak context, pass/N, AX

Short definitions, each with what it looked like in our own captures and runs, and a link to the page that goes deeper.

Leonid Bugaev · Published 21 Oct 2026

How agents load tools

tools/list

The MCP call that returns a server's tools with names, descriptions, input schemas and annotations. One 99-tool list we measured was 68,730 tokens.

Preloading

Sending every tool definition to the model with every request. OpenCode 1.18 did this: +73,415 tokens per request for 99 tools. See OpenCode 1 vs 2.

Deferred tools

Tools not sent up front. In Claude Code the model got a reminder listing 113 names only, then loaded definitions when it chose them. See Claude Code tool search.

Tool search

A built-in tool the model calls to find and load deferred tools: by keyword, or with select: and exact names. Grok Build uses BM25 keyword ranking.

Code mode

Codex's default: MCP tools sit behind one code-running tool, and the model writes JavaScript to filter ALL_TOOLS and call them. See how Codex finds MCP tools.

alwaysLoad

A Claude Code _meta flag that sends one tool's definition up front; it cut tokens before the first call from 15–28k to 6–7k in our test.

Discovery rate

Runs where the agent found and called your tools ÷ all runs. Codex: 34/40 in one live round.

Context and cost

Peak context

The largest single model request in a run: how full the window got. Codex averaged 18.1% of its 258,400-token window on our live jobs.

Cumulative input

All model requests in a run added up; roughly what's billed. About 210–218k tokens per simple question in our live runs. See MCP context window usage.

Context pollution

Tokens in tool results that didn't help: identical repeats, oversized errors, refusal boilerplate, the agent's own logged calls, unused records. See how much MCP output is waste.

Output limit

The point where an app cuts or files away a tool result: 25,000 tokens in Claude Code, about 10,000 printed tokens in Codex code mode, 51,200 bytes in OpenCode. See output limits by agent.

Cost per success

Input tokens over all runs of a job ÷ runs that passed. See cost per successful task.

Measurement

Job

A real user request in the user's words, without naming tools.

Ground truth

The right answer, from fixed tool calls made outside the agent, not from agent answers.

pass/N

Passing runs out of N. We never quote a single run.

Wilson range

A 95% confidence interval for a pass rate that behaves well at small N: 34/40 is 71–93%. See why pass rates need ranges.

Silent failure

A confident wrong answer, such as “that data isn't available” when it is.

Allowlist proxy

A proxy that forwards only named read tools to a production server, for safe testing. See test against production read-only.

Server design

Server instructions

Text an MCP server sends at connection. Apps cut it differently: 4,096 characters in Claude Code, about 2,048 in Grok Build. See server instructions length.

Annotations

Hints on each tool: readOnlyHint, destructiveHint, idempotentHint, openWorldHint. Codex exec refused destructive-marked tools without approval. See tool annotations.

Two-step confirm

A first call that changes nothing and returns the cost and a token; a second call with the token acts. See the two-step confirm pattern.

Next-step hint

A field in a result naming the next tool to call. See next-step hints.

Refusal state

What an error tells the agent to do: fix arguments, retry, or ask the user. See agent refusal states.

Agent experience (AX)

How well AI agents do your users' jobs with your product. See agent experience (AX).

Go deeper: Agent-ready tools, Test MCP servers with agents, How agents load MCP tools.