Turn skills on in the agent builder’s Skills tab.
What each skill gives
Section titled “What each skill gives”| Skill | What the agent gets |
|---|---|
| Browser control | Its own tab group to read, click and type in. Needs the AWFlow extension. |
| Web search | Searches the web in a background tab and reads the results. Needs the AWFlow extension. |
| Workflows | Each chosen workflow becomes a tool. The agent can pass it some text and gets back what the workflow produced (the output of its last steps). A workflow that pauses for an approval or a wait can’t finish inside the call, so the agent is told it failed. |
| On-device AI | Image, text and audio models that run on your device. See Local AI. |
| Teammates | Can ask the teammates you pick to do part of the work. Each works on its own model, skills and rules and sends back its answer. Teammates don’t see the chat, so the agent gives them a complete task. In a team thread, the team’s own handoff map is used instead. |
| MCP | The tools of each MCP server you add. For a server that needs a login, pick a saved credential (a Bearer token or API key from Credentials, see Connect any API) next to it; only a reference to the credential is stored on the agent. Each call follows the MCP rule, with a row per server on the Rules tab: Cautious and Balanced ask you first, Autonomous lets it act. Always allow applies to that server only. |
Skills need a model that can call tools. Chrome’s built-in Gemini Nano handles a few simple skills (up to 8 tools at a time, no browser control); with a chat-only model, such as some Ollama models, the agent’s skills are off. When a skill can’t be used (no extension, a deleted workflow, an MCP server that doesn’t answer), a banner above the composer says so, with a fix, and the agent says so instead of pretending. See When a skill is off.
Tool results and timeouts
Section titled “Tool results and timeouts”- What a tool brings back from outside (web pages, web search, MCP servers, workflows and teammates) is given to the model as data to read, never as instructions. The step shows Page text treated as data, and a warning when the text tried to instruct the agent. See Prompt injection.
- Each tool call has a time limit: 60 seconds, or 2 minutes for a workflow (a teammate has no limit). The clock pauses while the agent waits for your approval. A tool that runs out of time shows Timed out and the agent continues without it.
- Very long results are trimmed before they reach the model (the step says Result trimmed), and the agent can’t call exactly the same tool with the same input over and over: repeats are skipped.
Teammate results
Section titled “Teammate results”When an agent hands a task to a teammate, the answer comes back with a status (done, partly done when it ran out of steps or was stopped, blocked when an approval was declined or its model failed, or needs you when an approval wasn’t answered), so the agent knows whether to rely on it, retry or ask you. Ready-to-use pieces (a drafted email, a table, a list, links, code) come back as separate items the agent can use directly.