---
id: con-agents
schema: concept/1
layer: garden
status: established
name: Agents
aka: ["AI agents", "agent mode", "agentic AI", "computer use", "Cowork", "operator"]
summary: >
  AI that works toward a goal over several steps on its own, choosing actions, using tools and
  checking its progress, rather than giving one reply and waiting.
first_seen: "2023"
relations:
  builds_on: [con-tools-and-connectors, con-reasoning-models]
  evolved_from: []
  related: [con-steering, con-human-in-the-loop, con-automation, con-vibe-coding, con-privacy-and-data]
  parallels:
    - idea: "an executive assistant acting on your behalf"
      field: "administration"
      note: "You set the goal and the limits; they handle the steps and check in on anything unusual."
    - idea: "a general contractor"
      field: "construction"
implementations:
  - { provider: anthropic, name: "Claude Cowork", resource: res-anthropic-intro-cowork }
  - { provider: anthropic, name: "Claude Code" }
  - { provider: openai, name: "ChatGPT agent" }
  - { provider: google, name: "Gemini Agent" }
  - { provider: perplexity, name: "Perplexity Computer", resource: res-perplexity-search-vs-computer }
  - { provider: perplexity, name: "Comet browser" }
competencies: [cmp-steering, cmp-frame-the-outcome, cmp-use-responsibly]
origin:
  kind: ai-proposed
  by: "seed agent 2026-10-16"
  rationale: "The commons is built for a world of agents; this is where the user's role shifts from writer to director."
---

A chat assistant gives one answer and waits for you. An agent keeps working toward the goal. Given a goal like "collect the prices from these five vendor sites into a comparison table", it plans steps, opens pages, reads, fills in the table, notices a missing number and goes back for it, and then reports. It uses tools and connectors to act, and a reasoning model to decide what to do next.

With an agent, you spend less time doing the work yourself and more time directing it: stating the goal clearly, setting limits on what it may touch, approving risky steps, and checking the result. Agents can save hours on tedious multi-step work. They can also go confidently in the wrong direction for a long time, or do more than you asked.

Start with low-stakes, easy-to-check tasks, keep permissions narrow, and watch the first few runs.

## Example

An office manager asks an agent to read twelve supplier invoices in a folder, list each amount and due date in a spreadsheet, and flag any over budget. She spot-checks three against the PDFs before trusting the rest.

## How it connects

- Built from **tools and connectors** plus **reasoning models**.
- Working with one well is mostly **steering**, with a **human in the loop** for key decisions.
- Agents that run on a timer become **automation** or **scheduled tasks**.

## Try it

Pick a boring multi-step task. Before handing it over, write the goal, one thing the agent must not do, and how you'll check the result. Then try it in your tool's agent mode.