The Boston Builder Series, from @MGMT Boston in partnership with @Trove .
Pranav Sonje, AI Manager, Product Demand & AI Agents at @Blitzy, built two agent workflows that take a sales rep from a company name to a ranked account brief and four drafted emails in minutes. Reps trigger all of it by asking Claude.
Sales reps have never had more signals to work with: earnings calls, job postings, podcasts, LinkedIn posts. Sorting them is the slow part. Pranav's answer is Blitzy Scout, two agentic workflows with more than 300 nodes between them, each made up of multiple sub-agents.
By Blitzy's own tracking:
• 3,000+ accounts researched in HubSpot
• $100M+ in pipeline covered
• 70+ rep requests in the most recent week
1. Research Agentic Workflow: Who to call, and why
A rep names an account. The workflow pulls from five kinds of live sources: public filings, the latest earnings call, current job openings, coverage across the open web (podcasts, YouTube, LinkedIn), and the account's own HubSpot history.
Every signal is sorted into a theme, deduplicated and scored for confidence. Weak ones are dropped. What survives carries a short note on why it matters to Blitzy's pitch.
Then it ranks the account's contacts by seniority and fit, checks that ranking against the people actually on the account, and writes a short reason to reach out onto the top contacts' HubSpot records. Thirty to sixty minutes of digging becomes a few.
2. Outreach Agentic Workflow: Four emails, none sent automatically
The brief feeds a four-touch email sequence, drafted in one pass:
1. An opener built on the strongest signal.
2. A follow-up on a different signal, so it never repeats itself.
3. The pitch, matched to that contact's role.
4. A close that calls back to the opener.
Every draft is checked against Blitzy's approved messaging. If the workflow doesn't know enough about a contact, it flags them instead of guessing. Nothing sends on its own: a rep reads, tweaks and sends.
3. The MCP: it all runs from Claude
Pranav built a custom MCP that connects both workflows to Claude, so reps ask in plain English ("research this account") instead of opening another tool. The connector is a switchboard. It routes each request to the right workflow and follows standing rules, such as confirming something exists before describing it. Ask for a brief that hasn't been built and it says so.
The takeaway for operators
After 300 nodes, Pranav's lesson has little to do with scale:
"Autonomy without judgment isn't intelligence, it is just speed. The next decade will be teaching AI how to navigate responsibly."
That judgment is built into every layer of Blitzy Scout. Weak signals get dropped, unknown contacts get flagged, and a missing brief gets reported as missing. When the system does have an answer, he wants it precise and on point, not a bluff.
The Boston Builder Series highlights people across Boston building with AI, from @MGMT Boston, in partnership with @Trove.