Coding agents and the future of growth
Notes from a private dinner in New York City.
First, a little context.
We run a dinner series called The Growth Assembly. It all started two years ago when we saw the writing on the wall, that with writing code being more commoditized, distribution was going to become more valuable than ever before.
Our (very selfish) motivation was to bring together the most cutting-edge growth operators for a private dinner and have them share what they're working on, what they're curious about, and their predictions for the future.
The last dinner we hosted was the most significant one yet.
We invited 10 of the top growth leads across NYC startups (Series A+) to learn more about how they were using coding agents within the GTM org.
There was only one caveat: to attend the dinner, everyone had to present their favourite or most impactful coding agent workflow. We found (with much difficulty) a private room in a steakhouse in midtown with a screen. Everyone brought their laptops and took turns demoing their workflow to the room and then fielding questions.
We honestly had some apprehensions about whether the format would work or if this would be a huge waste of everyone's time.
This turned out to be the most insightful couple of hours we've had in recent times. It was both a glimpse into the future and a huge reality check on how much the growth function is evolving and iterating at a pace that's hard to comprehend.
We had multiple Granolas running (with consent) and what follows is what we learnt.
Before the demos: a few observations
Shift to central agents. A big shift is happening from a few people in the org having very powerful but individual coding agent setups to one central agent that everyone in the org can access. A lot of people working in growth may never have to go into the nitty gritties of setting up their own coding agent stack while still getting the efficiency gains AI gives.
So much of what was cutting edge 6 months ago is table stakes now. For example: accessing large amounts of data across CRM and GTM tooling, skills for recurring tasks, building sales and marketing material on demand, etc.
Shift to player-coach. A decent portion of the room lead the function, have several direct reports, but are very close to bare metal and doing IC work. Elena Verna has a really good piece of writing about the 'High Impact Individual Contributor.'
Earlier the company, the faster their adoption. So much of what these companies were able to achieve was because of how much autonomy they had to work with internal data and pipe tools in. Half of the effort in building these for a larger org would be to get approval from compliance, legal, etc.
The demos
A quick note: we follow Chatham House Rules. What follows are the insights and workflows, but not the names or companies behind them.
-
+The junior employee that lives in Slack
What they built
"Theo," an open-source coding agent that the team has set up to operate as a long-running junior employee. The team talks to him in Slack the same way they'd talk to any teammate, and he picks up tasks across their entire go-to-market stack.
How it works
Theo runs on a persistent box (a Mac Mini sitting in the corner of their office) and has his own identity in the company Slack. Under the hood, a few pieces work together:
- A Slack interface layer that lets him be tagged or pinged in any channel. He responds inline, holds context across a conversation, and remembers what was said earlier in the thread.
- API and MCP connections to their core tools, including Meta ads, Google ads, their analytics stack, and a handful of internal systems. Access is read-only on anything that could break the business.
- A skills layer where the team has written down how Theo should approach different kinds of asks: competitive research, ad-account analysis, "before you do X, check Y," and so on.
- Persistent memory and feedback. When he gets something wrong, the team corrects him and writes the correction down, and he's better at it the next time.
Common use cases
- Fixing broken Excel formulas without pulling an ops or finance person in
- Running ad-hoc competitive scans (the founder recently had him pull TikTok Shop competitor pages into Figma)
- Standing up small dashboards nobody wants to spend a sprint on
- Monitoring metrics and accounts that would otherwise need a person actively watching them
- In one case, ordering himself a phone so he could install and test mobile apps directly
Caveat: This was a meaningfully bigger technical lift than most of the other workflows in the room and needed to be built with the engineering team.
-
+The ROI calculator to help close more deals
What they built
A public ROI calculator where a prospect enters their company domain and gets back a personalized PDF report estimating what their data-quality problem is actually costing them. Each prospect gets a different report.
How it works
The prospect goes to a page on the website, types in their domain, and hits generate. A coding agent runs queries about the company (engineering headcount, infrastructure scale, public information about their data stack) and uses it to build a profile. It plugs those inputs into the company's internal data-quality cost framework.
The report shows up as an editable web page. If the prospect thinks the agent got it wrong (say they actually have 100 engineers instead of 85), they can change the number and the rest of the report updates immediately. A download button spits out a clean PDF the prospect can forward to their CTO or CFO.
Why it works
Most sales follow-ups go out as generic assets like a deck, a one-pager, or a case study, and none of them do much to actually move a deal forward. An ROI report is more compelling. And because the prospect can edit the math themselves, it doesn't behave like a static PDF.
-
+Hyper personalized LinkedIn Ads
What they built
A paid ads engine that generates personalized LinkedIn ad creative for individual target companies and launches the campaigns programmatically. Each target account gets its own dedicated ad with messaging written specifically for them. The first version was built in a single night.
How it works
- Audience. Target accounts pulled from their internal go-to-market data, which they happen to sell as a product, so the audience is already rich.
- Creative. They use Paper, a design tool whose MCP outputs raw React and HTML rather than pixel-based files. Because the output is code, Codex can manipulate it directly. The marketer tells Codex what each target company cares about, and Codex generates ads in every dimension needed.
- Launch. LinkedIn's campaign manager requires a minimum audience of 300 per campaign, making 1:1 ABM impossible through the UI. The LinkedIn Ads API doesn't have that restriction. Codex launches campaigns directly with names, audiences, UTMs, and budgets all configured correctly.
- Tracking. A second Codex automation pulls performance stats back into the workflow so the marketer can iterate without going back into LinkedIn's UI.
In one sitting, this generated and launched 36 ads against three target accounts. It scales cleanly to a thousand.
-
+A BDR workflow for every inbound demo request
What they built
A workflow that takes every demo signup on the company's website, qualifies it overnight, drafts a personalized outreach email for the account owner, and pushes the draft into Outreach.io ready to send. The AE only has to review and click.
How it works
The pipeline runs on Claude's scheduled tasks plus a GitHub Actions job. A demo-page visit records the lead in the CRM and an enrichment step (running through Apollo's MCP) attaches firmographic data. A Claude scheduled task runs once a day, pulls that day's demo requesters, and qualifies them. Anyone who clears the bar gets handed off to a chain of five Claude skills that combine to produce one personalized outreach email.
A GitHub Actions job takes the drafted email and pushes it into Outreach via the API, then assigns a task to the account owner saying "this email is ready, review and send."
-
+Customer data on demand for the entire GTM team
What they built
A Claude skill called "Lookout" that gives anyone on the GTM team a self-serve natural-language interface to customer data. Anyone can ask "what happened with [customer X]?" and get a real-time pull of every login, click, and feature touch, without going through engineering or waiting for a dashboard to refresh.
How it works
An engineer built a read-only replica of the production database and routed it through AWS. The skill (Lookout) is what the team actually interacts with. When someone asks a question in Claude, the skill runs the relevant query against the read-only mirror and returns a structured answer.
Caveat: Similar to one of the earlier demos, this also required real engineering lift. It's a signal that more go-to-market teams should be working directly with their engineering teams to build these more complex workflows.
-
+An in-house AI SDR for SMB outbound
What they built
An AI SDR built in-house in Claude Code, after trying a handful of off-the-shelf AI SDR products first and finding that none of them could personalize well enough for the kind of SMB outbound they needed to run at scale.
How it works
Leads come from Data Axle, an SMB-specific data provider that scrapes Google Maps and other sources. For each lead, the workflow visits the business's website and pulls out additional attributes. The agent then runs the lead through four skills in sequence:
- ICP context. What their ICP looks like, what messaging has worked in past campaigns, what's distinctive about the kind of business this is.
- Personalization angle. Scores the lead's attributes and decides which angles to take for this specific account.
- Cold email best practices. Pulls from a curated set of frameworks (Josh Braun and others) to shape the structure.
- Sequence architecture. Assembles the personalized email sequence and outputs the final copy.
The drafted sequence gets pushed into SmartLead through an MCP connection. Alongside the email, the agent generates a personalized audit of the prospect's existing booking site, showing where the gaps are.
The first version is converting 3-5x better than the off-the-shelf AI SDR products they tested.
-
+A TikTok scriptwriter that runs every morning
What he built
A daily workflow that scrapes 100+ sources every morning, ranks the most viral content of the day, and writes three TikTok scripts ready to record. Andrew records and edits them on his phone in about two minutes per video. The system took his TikTok from 71 to roughly 70,000 followers over 30 days.
How it works
- Scrape. A scheduled script pulls fresh content from 100+ sources, including independent newsletters, targeted subreddits, and Substacks hand-picked as good signals for emerging viral topics.
- Rank. Each scraped item gets scored against a custom virality rubric grounded in Jonah Berger's Contagious (eight factors that drive viral spread) and a handful of essays on consumer social. The system pulls the top 25 most viral items.
- Write scripts. The top three items each get turned into a TikTok script using a separate script-writing framework. The prompt asks Claude to "think like Jonah Berger, think like Nikita Bier."
- Output. Each day's three scripts land in a Google Doc, and a Notion MCP keeps a lightweight project tracker.
Out of the three scripts each day, at least one consistently hits 10K+ views, often much more.
-
+Using your own voice AI to scout prospects
What they built
A workflow that uses the company's own voice-AI product to call a prospect's actual phone lines and build a dataset on how well that prospect is handling inbound customer calls. The SDR then walks into the cold call with hard, specific numbers about the prospect's real customer experience.
How it works
For a multi-location prospect (say, a 200-location HVAC chain), the team points their voice-AI at the prospect's published phone numbers and runs a batch of test calls across different times of day. The system pulls back call answer rate, quality of the human handoff, average hold time, and conversion behavior.
The output is a one-page sales brief the SDR can lead with: "I've been studying your customer experience across all 200 locations for the last week. Here's what we found."
Why it works
The hardest part of selling to a large multi-location business is getting them to take the first conversation seriously. A voice-AI listening test is a piece of data the prospect doesn't have on their own business, so the conversation immediately starts from a place of real insight. The company is also using their own product to do the work, which both demonstrates capability and costs them almost nothing per call.
-
+A negotiation skill built from three opposing personas
What they built
A Claude project they call "Three Will for Sales" (named after the company's first salesperson), where every outbound message in a live negotiation has to be approved by three different Claude personas before it gets drafted. Each persona looks at the negotiation through a different lens, and the email doesn't ship until at least two of them agree.
How it works
When a negotiation comes in, the operator drops the latest email exchange and the relevant contract info. The system then runs the situation past three personas:
- Logical Will. Looks at the deal from a pure trade-off and incentives standpoint. Are the numbers fair? Are the terms aligned?
- Emotional Will. Thinks about how the message will land on the other side. Will the person feel respected? Will they get defensive?
- Communication Will. Stress-tests how clearly the logic and emotion come across in the actual language.
Each persona evaluates independently, and they actually disagree with each other, sometimes loudly. The system won't draft the final reply until two of the three sign off. A separate skill rewrites the final draft in the operator's own voice, using their sent-email history as a style reference.
The whole thing also runs as a scheduled task at 5pm every day: any negotiation or contract email from the day gets auto-drafted and is ready for review.
About us
