The first agents run inside Workbench

December 2025 in Workbench: the discovery report became a Workbench deliverable with its own crawler and local scans, the first agents were defined and run under a staff member's eye, planning agents began writing plan items, and a plugin connected client sites.

December 2025 is the month the discovery report became a thing Workbench produces end to end, and the month the first agents ran. The work that later became visible to partners started here, out of sight.

For our partners

Two things reached partners. From December 7 a discovery produced in Workbench is sent back to the Portal and shared as a report, with a slide deck and a presentation view built from the same data, and from December 18 the report includes a competitive matrix. From December 2 a proposal can be shared by link, so a partner reads the current version rather than an attachment.

Behind the scenes

1. The first agents were defined on December 4 and 5. Two to begin with: a campaign planner, which reads what we know about a partner’s client and drafts a campaign plan, and a content planner, which turns keyword research and campaign goals into a content plan.
2. The discovery deliverable was built end to end. By December 11 a discovery could be worked start to finish in Workbench: request in from the Portal, crawl, keywords, local scan, competitive review, summary, proposal, share link out.
3. On December 23 two planning agents, one for links and one for SEO work, were given the ability to create, update and schedule plan items directly. A plan item is what becomes a task, so this was the first time an agent’s output went into the same queue our specialists work from, with the same lifecycle and the same notes.
4. On December 20 Workbench got real-time notifications and an internal chat, and on December 22 you could see who else was looking at the same page.
5. On December 24 a small plugin for client sites let a site be linked to its campaign in Workbench, and keywords, competitors and locations per campaign followed between December 27 and 31. Both were groundwork for agents that act on a site rather than only write about it.

1. The first agents

Each agent has a description of its job, a fixed set of things it is allowed to do, and a chat where a staff member watches the run, reads the output and sends it back with feedback. Agents keep their own run history and can read the feedback on earlier runs, so a correction is not lost the next time.

A content researcher followed on December 9, and a discovery agent on December 11 and 14 that reads the crawl, keywords and scan data and writes the summary and proposal sections a person then edits. On December 13 agents became instances: one per partner, campaign or employee, each with a record of its runs, first-time approvals and rounds of feedback. We wanted the agents measurable from the start, not after.

The agents in December 2025 were watched constantly, corrected often, and used by a handful of people. The pattern we still use was set then: an agent drafts, a person reviews, the review feeds the next run, and the results are counted.

2. The discovery deliverable

On December 3 Workbench got its own site crawler, and the discovery gained an audit built from the crawl, a local section built from the prospect’s Business Profile and local rank scans, and a competitive section built from nearby businesses in the same category. A proposal editor with a live document view had arrived on December 1.