Agents that work for your brand: what we added since August and why
In four weeks, Branddock got a full agent platform: build your own agents, rooms and teams, agents that respond to events, and per-tool access that you control. Here's what you can do with it, and what we're deliberately not doing yet.
A marketer has more work than hours. Research, briefs, reviews, reporting: these are tasks that keep coming back, and they call for exactly the knowledge of your brand that's already in Branddock. Since the end of August you can hand that work to agents who bring that knowledge with them: each with its own role and a track record you can review per run.
Your own agents, in five steps. The wizard starts with a starting point (an example from the library or a blank one), asks for a persona and instructions, lets you review the result and, if you want, puts a schedule underneath it. You don't need to know anything about prompts: your Workspace's brand DNA, voice and personas come along automatically. Every custom agent has a monthly credit cap, so an agent you forget about can't keep running unnoticed.
Rooms, teams and meetings. In a room you work with agents the way you work with colleagues in a chat: you address them by name, they respond to what's on the table. You reach teams with @team; a team meeting produces minutes and proposals. Every action by an agent is a proposal. You approve it, even straight from the email you get about it, or you leave it as a draft. An agent that publishes on its own doesn't exist in Branddock; that's by design, not a limitation of the first version.
Access per tool, per agent. Recently every agent has a Tools and access section. There you decide whether it may go on the web, and per connected tool whether it may only read or also act. An agent with permission to act on your self-hosted WordPress prepares a draft for you and nothing more: no publishing, no changes to existing pieces. Without access it doesn't even see the tool. "What can this agent do" is thereby a setting you can review, not an assumption in a prompt.
Responding to what happens. Until now an agent ran on a schedule or on your button. Now it can also start as soon as someone requests a review, as soon as a piece goes live, or as soon as a connected tool finishes syncing. Events that come in within five minutes are bundled into one run; there's a cap per event, per agent and per day. Such a run costs three extra credits, on top of what the agent itself charges. An agent's autonomy setting stays on "Always ask for approval" until the emergency stop is there, with which you break off a running run right away; that emergency stop is coming soon.
What we're deliberately not promising yet. We're working on a catalog of skills: small, reusable instruction sets that make an agent better at one task, like writing an SEO brief or reviewing a press release. We measure every skill first: two runs with, two runs without, on a fixed test task. Anything that makes no measurable difference doesn't make it into the catalog. The first measurements are running now; as soon as the first batch passes, you'll read it here. Until then: everything an agent does today, you can review in the app in its runs, with every step included.