How do the top AI agent harnesses handle tool calls?
The seven harnesses ship the same core tools (read, write, edit, run a command) and differ on who approves a call. OpenCode and pi run tools without asking by default; Antigravity's default asks only outside a sandbox; Codex sets a sandbox and a separate approval policy; Claude Code layers allow, ask and deny rules that hooks can tighten. Read 2026-11-06.
Contents
Short answer, from the docs read: the tool list is nearly the same everywhere, and the real difference is the default for who approves a call. pi and OpenCode run tools without asking unless the user turns on checks.[1][2] Antigravity's default lets commands run inside a sandbox and asks before leaving it.[3] Codex splits the question in two: a sandbox sets what a command can touch, and an approval policy sets when the agent must stop and ask.[4] Claude Code lets a user write allow, ask and deny rules per tool, and its docs say a deny rule keeps a call refused and a hook that blocks a call stays blocked.[5] Cursor's agent docs list its tools but the approval settings were not read.[6] This is the third page in a series that began with a survey of the harnesses and continued with memory.
A tool call is the moment the model asks the harness to do something, such as read a file or run a command, and the harness decides whether to do it and returns the result.
The tools they ship
- pi ships four: read, write, edit and bash. Read-only tools (grep, find, ls) exist but are off by default. Its author argues four are enough, because models are trained on bash and on read, write and edit tools with similar inputs.[1]
- Cursor lists file editing, codebase search, web search, terminal execution and browser control as Agent's tools, and says there is no limit on the number of tool calls in one task.[6]
- Hermes Agent describes a broad built-in registry covering web search, browser automation, terminal execution, file editing, memory, delegation and scheduled tasks, organized into toolsets that can be enabled or disabled per platform. Plugins add more.[7]
- OpenCode comes with a set of built-in tools, and adds custom tools and MCP servers.[2]
- Claude Code documents a long tool reference with per-tool behavior, including a Bash tool, a Grep tool, notebook editing and an Agent tool that launches subagents.[8]
- Codex and Antigravity: the pages read describe how calls are approved but did not give a tool list.
Who approves a call
- OpenCode. By default all tools are enabled and need no permission. A permission setting resolves each tool to allow, ask or deny, with wildcards for groups such as every tool from one MCP server. Starting with --auto approves every request that is not explicitly denied.[2][9]
- pi. Full "YOLO mode": no permission prompts for file operations or commands. The author's reasoning is that a tool which can read data, run code and reach the network cannot be made safe by prompts, so the harness gives in.[1]
- Antigravity. Four presets. Request Review asks before every terminal command. Default runs commands inside a terminal sandbox without prompting and asks for anything outside it. Auto lets Antigravity evaluate commands that need to leave the sandbox. Turbo runs everything with no restrictions. Allow, deny and ask rules the user sets sit on top of the preset and take precedence.[3]
- Codex. Sandbox modes are read-only, workspace-write (the default low-friction mode for local work) and danger-full-access. Approvals are a separate setting: the docs say the sandbox defines technical boundaries and the approval policy decides when the agent must stop and ask. A desktop option, Approve for me, sends eligible approval requests to an automatic review instead of a prompt.[4]
- Claude Code. Allow rules use a tool without asking, ask rules prompt every time and deny rules prevent use. The docs list modes including acceptEdits, auto, plan and dontAsk. The docs' section on precedence among rules, hooks and mods says a call a deny rule refuses stays refused (an organization can change that), a call an ask rule would prompt for still prompts, and a call a PreToolUse hook blocked stays blocked whichever settings file the hook comes from.[5]
- Hermes Agent. The docs read cover where commands run: a terminal backend that can be local, Docker, SSH, Modal, Daytona or a Vercel sandbox, among others.[7] The approval settings were not read.
- Cursor. Not read.
Where the line between sandbox and permission falls
Two harnesses treat a sandbox as the thing that makes asking unnecessary. Antigravity's Default and Auto presets run commands without prompts only inside a sandbox.[3] Codex says the sandbox reduces approval fatigue, because the agent can run routine commands inside a boundary the user already approved.[4] Hermes Agent offers the sandboxed backends but the page read does not say whether it prompts.[7]
What the comparison rules out
These are consequences of the docs read, not findings from a test.
- "Does the agent ask first?" has no single answer per harness, because each has settings that change it. The defaults are the useful comparison: OpenCode and pi do not ask; Antigravity asks outside its sandbox.
- A harness that does not ask is not necessarily unsafe, and one that does ask is not necessarily safe. Claude Code's docs themselves say that a CLAUDE.md instruction is guidance, not a boundary, and that a deny rule or a hook is the enforcing layer.[5][10]
- Counting tools does not separate the harnesses. pi ships four and says that is enough; Claude Code's reference is far longer. Whether more tools help was not tested in any source read.
Terms
- Tool call: a request from the model to run one of the harness's tools, with inputs.
- Sandbox: a boundary that limits which files and network a command can reach.
- Approval policy: the rule for when the harness stops and asks a person before running a call.
- Hook: a script the harness runs at a fixed point, such as before a tool call; the next page covers them.
Not yet proven
- Safety. The docs describe controls, not how well they hold. No source read tests whether a sandbox or a deny rule stops a real attack.
- Cursor and Hermes Agent approval settings, and Codex's and Antigravity's tool lists. Not read.
- Defaults for Claude Code. Its default mode was not read, so this page does not say what a fresh install does.
- Which tools models use best. pi's author claims four tools are enough; no benchmark was read.
- Currency. Antigravity, Codex and Claude Code have changed their permission settings by version; these pages were read on 2026-11-06 and may differ on an older install.
If you run one of these harnesses and its defaults surprised you, I would like to hear it: mackenzie@mackenziebowes.com.
Sources
- Mario Zechner, "What I learned building an opinionated and minimal coding agent", 2025-11-30. Read in Rotunda 2026-11-06. mariozechner.
at/ posts/ 2025- 11- 30- pi- coding- agent/ - OpenCode docs, Tools. Read in Rotunda 2026-11-06. opencode.
ai/ docs/ tools/ - Google Antigravity Docs, Agent settings. Read in Rotunda 2026-11-06. antigravity.
google/ docs/ agent- settings - OpenAI, Codex docs, Sandbox. Read in Rotunda 2026-11-06. learn.
chatgpt. com/ docs/ sandboxing - Claude Code Docs, Permissions. Read in Rotunda 2026-11-06. code.
claude. com/ docs/ en/ permissions - Cursor Docs, Agent overview. Read in Rotunda 2026-11-06. cursor.
com/ docs/ agent/ overview - Hermes Agent docs, Tools and Toolsets. Read in Rotunda 2026-11-06. hermes-
agent. nousresearch. com/ docs/ user- guide/ features/ tools - Claude Code Docs, Tools reference. Read in Rotunda 2026-11-06. code.
claude. com/ docs/ en/ tools- reference - OpenCode docs, Permissions. Read in Rotunda 2026-11-06. opencode.
ai/ docs/ permissions/ - Claude Code Docs, Memory. Saved 2026-10-22. code.
claude. com/ docs/ en/ memory
Terms on this page
- Hook: A command that an AI coding agent runs automatically at a set point in its work.
- Model Context Protocol: An open standard for connecting AI applications to outside tools and data.
- Tool call: A request from an AI model to run one action, such as a search or a file write, which the surrounding program carries out.