AI agents
Google and OpenAI now sell “AI coworkers.” Do you still need a custom agent?
Google’s Gemini agent and OpenAI’s Dots bring general-purpose AI agents to work. Where they fit, where a custom agent still wins, and why MCP matters.
In the space of ten days, the two biggest names in AI announced agents meant to work alongside your staff. Not chatbots that answer questions, but software that takes a goal, works through the steps, and keeps going after you close your laptop.
If you run a small business, you’ll hear a lot about these. Here’s what was announced, where general-purpose agents fit, and where an agent built around your business still makes more sense.
What happened
OpenAI’s Dots (September 29). At its DevDay event, OpenAI introduced Dots, which it calls “remarkably capable, always-on agents.” Each runs on GPT-6 Astra, OpenAI’s flagship model, and can connect to more than 4,000 apps. The first Dot is included with the Pro and Business Premium plans in eligible markets. OpenAI’s developer platform also gained “computer use”: an AI that clicks, types, and finds its way around software the way a person would.
Google’s Gemini agent (October 8). Google announced an agent that takes “objectives, not just instructions,” in the words of Google Cloud CEO Thomas Kurian. It gets its own Workspace account and email address, and it can connect to Google Workspace, Microsoft 365, Slack, Jira, and databases such as BigQuery, Postgres, and Snowflake, as well as any server that speaks the Model Context Protocol. Businesses get it first. Google didn’t announce prices, but it described spending options that include real-time spend caps.
The plumbing got an upgrade, too (July 28). The Model Context Protocol, or MCP, is the open standard that lets AI models connect to business tools. Its biggest update yet makes connections simpler to host, formalizes how extensions work, and tightens how AI apps get permission to connect. The official MCP software kits now see close to half a billion downloads a month.
Where general-purpose agents fit
If your business already runs on Google Workspace, Microsoft 365, or ChatGPT’s business plans, these agents are worth trying for office work:
- sorting and drafting email;
- scheduling across calendars;
- summarizing documents and meetings;
- pulling numbers together from spreadsheets and shared drives.
They’re general by design. They know how to use common apps, and they’ll keep getting better at it.
Where a custom agent still wins
Your customers. General agents work inside your office suite. The conversations that win or lose business, on your phone line, your website, and your intake forms, need your rules, your tone, and a clean handoff to your staff.
Your systems. Plenty of small businesses run on software that isn’t among the 4,000 apps: practice management, dispatch, inventory, an industry database, or a spreadsheet that runs the whole operation. A custom agent is built around the systems you actually use.
Your rules. A custom agent can be held to exactly what it’s allowed to see and do, with approval steps for anything sensitive and a log of every action. That matters more as agents gain access to email and files. (See our post on shadow AI and safer agents.)
Your data and your costs. If customer data shouldn’t leave your building, an agent can run on your own hardware, at a fixed cost instead of a meter.
Why MCP matters to you
MCP is the reason this isn’t an either-or choice. If the connection to your scheduling system or database is built as an MCP server, the same connection can serve Gemini, ChatGPT, Claude, or a model running in your office. You build it once, and you aren’t locked in to one vendor’s agent, prices, or roadmap.
There’s a cost. The new specification admits “some migration cost” for existing connections, and some older features are being phased out over at least twelve months. For new work, building on the current version from the start avoids that.
Our take
Use what you already pay for. If your team lives in Google Workspace or ChatGPT, try the built-in agents on everyday office work. Then put custom work where it pays: the customer conversations, the systems unique to your business, and the data that should stay private. Build those connections on open standards, so they’ll work with whichever AI you use next year.
That’s how we approach custom AI agents.
Sources
- Decrypt: OpenAI gave AI agents their own computers: everything announced at DevDay 2026, September 29, 2026
- TechCrunch: Google brings agentic AI to Gemini, starting with businesses, October 8, 2026
- Model Context Protocol: The 2026-07-28 Specification, July 28, 2026