Agent action

Definition

An agent action is a request from an AI coding agent for a tool to do something on its behalf, such as starting another chat or running a job, rather than only reading information.

Also called: agent request, tool action.

Coding agents such as Claude Code and Codex call tools, often through an MCP server. Many tools only read (list files, fetch an issue); an action changes something or starts work. Because an agent's model can be steered by what it reads, a prompt injection in a file for example, actions are where a person usually wants a say.

What makes actions safe to allow

• Approval outside the agent: the decision is made by a person or a rule, never by anything the agent sends.
• No escalation: whatever the action starts gets no more access than the agent that asked.
• Depth limits: agents started by agents can ask for less, so a chain can't grow by itself.
• Budgets and a kill switch: a cap on how many and how often, and one switch that stops them all.
• A log: every request, decision and outcome can be read afterwards.

In Pairbar

With Pairbar's tools installed, a Claude or Codex chat can ask for a second opinion or a review from the other app, a helper chat, a handoff, a follow-up to its own helpers, a notification, a workflow run or draft, or a check-in on itself. Each request gets an approval ticket and runs only if your settings allow it or you approve it in Pairbar; an agent can never approve anything. Guardrails are checked in order: paused or off, depth, no escalation, recent declines, repeats, budgets and the plan floor. See approve agent actions.