One agent holds the keys

Achal runs five agents: Muse, Instinct, GrokBot, Tiny, and Dots. Dots is the primary agent and the only one connected to his accounts, while the other four can see the Trello board and nothing else. Dots posts tasks to the board, the other agents pick them up and post their results back, and Dots reviews that work before bringing anything to Achal. He doesn’t check the board himself. He talks to Dots, and Dots comes to him whenever something needs a decision.

“I still want to control what I send.”

—Achal Aggarwal, CPO, Born West

‍

Keeping access with one agent makes the whole setup much easier to track. Google Workspace holds a lot of Achal’s information, including his email, documents, and calendar. If all five agents could reach it, he’d have no reliable way of knowing which one had touched what. With a single agent holding that access, there’s one clear place to look if something goes wrong.

“I've tried to keep a sensible structure, almost like a team with a hierarchy. Not a free-for-all where all of them get access to all of my data and I'm completely lost.” Achal Aggarwal

The web research runs separately on Muse and Grok, which each have their own cloud computer where Achal isn’t logged in to any of his accounts. If an agent clicks on something it shouldn’t while it’s browsing, that happens on a machine that holds none of his personal data.

‍

Finding companies the week they raise

The first workflow looks for companies that have just raised a Series A or Series B funding. It runs daily, and the cadence can be changed whenever Achal wants. For each company, the agents gather its market, team size, and tech stack, and then draft a message to whichever executive is most relevant, whether that’s the CEO, the CTO, or the chief product officer. Dots then checks every draft against Achal’s voice guide.

Each finished draft lands on Achal’s calendar as an event. The invite includes the message itself, why the company is relevant, the surrounding context, and the best channel and time to send it, whether that’s a LinkedIn message or an email.

“The calendar invite has all of the information I need: the actual message, the relevance, the context, why I’m sending it, and when.”‍
—Achal, CPO

When the event comes up, Achal reads the draft in the invite with the person’s LinkedIn profile open beside it and edits anything that needs changing. So far the drafts have needed minimal editing, and the first batch of messages goes out this week.

‍

An RFP agent that reads the fine print first

The RFP workflow follows the same structure. The agents search for RFPs across government and the private sector, collect the documents, and write a first-draft response based on what they already know about Achal and Born West. Dots reviews each draft and logs it in a shared spreadsheet with a status, so Achal and Rishav can see the entire pipeline in one place.

The most valuable part of this workflow happens before any drafting begins. RFPs often bury requirements that can rule a bidder out completely, and the agent flags those as early as possible. 

Catching these up front means we know where we stand on every requirement before putting a full day into a response.

“It’s already done the first draft, based on what it knows about me and about Born West.”‍
—Achal, CPO

Most of the RFPs the agents find are for migration and platform work, such as Drupal and WordPress sites or e-commerce builds moving from one system to another, and very few ask for something to be built from scratch. That gives us a useful picture of where public RFPs sit compared with the product work we usually lead with. Achal’s goal is to have five strong submissions ready by the end of October.

‍

Why nothing gets sent automatically

Dots is capable of doing a lot more than it’s currently allowed to. It has access to Achal’s browser, where he’s logged in to LinkedIn, so it could scan his connections, notice who has changed jobs, and message them directly. Achal hasn’t asked it to do any of that, and he doesn’t plan to.

His main concern is what the agents might send or post without his approval. A poorly written email is embarrassing, and a bad post from a company's LinkedIn page can spread widely before anyone has a chance to catch it.

He also doesn’t think a written instruction like “always ask before posting” would be enough to prevent that, because models aren’t deterministic and different instructions can end up pulling against each other.

“If I say ‘make me more efficient’ and also ‘don’t write for me’, it’s going to find a loophole in my instructions.”
—Achal, CPO

For that reason, the limits are set in the permissions themselves. Gmail access is read-only, Dots can write to Google Drive and nowhere else, and even in Drive it can’t share files with anyone except Achal. The other four agents have no Workspace access at all.

Achal also pays close attention to how an agent phrases its actions. He wants to hear “Can I make this change for you?”, and he’s uncomfortable when an agent says “I’ve already made this change,” even if the change itself was a good one. We wrote more about this in our September piece on agent permissions.

‍

Where it stands

The setup is still early. The outreach workflow has drafted and scheduled its first batch of messages, and the RFP workflow has first drafts and a shortlist in progress. Neither has brought in new business yet, and until it does, it’s a promising setup that still has to prove itself.

Achal’s case for it is about capacity. He couldn’t keep this research current every day on his own, and the reading, sorting, and fine-print checking that the agents now handle used to take up most of his week.

‍

Lessons you can steal

1. Give one agent access to your accounts and have the rest work through a shared board, so there’s only one place to look if something goes wrong.
2. Run web research on cloud computers where you aren’t logged in to any of your accounts.
3. Review each draft, then add it to your calendar with the context attached and post it when the timing is right.
4. Ask your RFP agent to flag disqualifiers like compliance, data residency, and staffing rules before it starts drafting.
5. Make email read-only and give the agent one place where it’s allowed to write.
6. Pay attention to whether an agent asks before it acts, even when the action would have been a good one.
   ‍

The agents can read as much as they need to, and the send button stays with Achal. We built this setup for ourselves in a few weeks.
If you want something like it for your sales or partnerships team, book a 15-minute intro and we can sketch out a first version together.