browser-use/agency
Agency card feed: agents finish work, Magnus decides with one click
What it solves
Agency is a system designed to help coding agents find and prepare useful work for humans to review. Instead of waiting for explicit instructions, it enables an agent to proactively research sources (like Gmail, Slack, and repositories), identify problems or opportunities, and prepare concrete fixes, replies, or ideas in the form of "cards" for the user to approve or reject.
How it works
Agency combines a local web application (the "host") with a set of instructions (the "skill") that a coding agent follows. The agent uses the skill to learn the user's goals and preferences from a private profile (me.md), researches data via authorized connectors and browser sessions, and then uses an API to push results as HTML cards to the local app. These cards present the problem, a proposed outcome, and a specific action (e.g., "Send" or "Apply"). The user reviews these cards, provides feedback, or approves the action, which the agent then executes and verifies.
Who it’s for
It is built for users who employ coding agents and want those agents to transition from reactive task-execution to proactive assistance, managing a feed of suggested improvements and communications.
Highlights
- Proactive Work Discovery: Automatically scans sources like Slack, Gmail, and Linear to find issues or outreach opportunities.
- Card-Based Interface: Presents findings as visual cards with screenshots, diffs, or demos, reducing the user's cognitive load to a simple decision.
- Customizable Layouts: Users can tell the agent to change the appearance of the cards (e.g., "Make them look like Pokémon cards") by editing a layout file.
- Private Profiling: Maintains a
me.mdfile to store the user's "dream," writing style, and constraints locally. - Human-in-the-Loop: Requires explicit human approval before any message is sent or code is merged.
Related
- Project
- Project
- Project
- Project