Win with Growth Loops
in an Agentic World
Your next thousand users will likely not be people, and most growth models don’t account for that yet.
Head of Growth at Amplitude
Sept 16, 2026
Who am I
Tommy Keeley is Head of Growth at Amplitude, where his team drives AI adoption & builds agentic capabilities into products like Slack & Teams. Before Amplitude he spent nearly four years driving growth at Dropbox, including Principal PM for Sharing & Virality & interim Head of Growth. Earlier he worked at high growth startups including Attentive & AlphaSense. He writes Tommy’s Thoughts on Substack.
Agenda
- 1The year of agents.. Slack, then Buzz
- 2A quick poll
- 3The 5-question audit.. what still applies
- 4What changed.. density, learning, cost
- 5Live audit.. your loop, scored
- 6Q&A
The hook
Slack & the year of agents
Each step was bigger than the last. Engagement loops & network effects used to mean human to human, but now they can be human to agent, agent to human, and even agent to agent.
- Late 2025Agent apps in channelsChat with Cursor like a coworker. The bots answer with 👀 and ✅.
- JanuaryAmplitude ships its agentUsers get answers & take actions on their data from Slack.
- AprilSalesforce Headless 360The whole platform via APIs & MCP. Slack is the front door for humans & agents.
- JuneClaude chimes in uninvitedReads the channel, jumps into threads, slinks away when told to pipe down.
- July 21Jack Dorsey launches BuzzA groupchat platform for teams of people & agents, built to replace Slack & GitHub in one shot.
The hook
Slack’s Buzz-kill
Two bets on the same future. Both products bet on ‘users’ becoming a blend of humans & agents over time.
| Slack | Buzz | The delta |
|---|---|---|
| Added agents to a human product | Agents as founding members | Who the product was built for |
| Agents live in Salesforce’s cloud, on Salesforce’s terms | Self-hosted & decentralized. You own the room & the agents in it | Who sets the rules of the loop |
| Still mostly conversation | Wires the chat to your repo. Talking & shipping in one place | Where the loop closes |
A quick poll
Do agents show up as a user type in your growth model today?
Answer in the Zoom poll. Thirty seconds.
The 5-question audit
A growth loop is still a growth loop
The boring old playbook questions still apply & are the right ones to ask. A node on the network takes an action, the action creates value, the value pulls in the next node, & repeat.
The 5-question audit
Node › Action › Compound › Rules › Speed
Who’s a potential node? A user, an admin, a team, an agent?
A user, an admin, a team. Each one scarce, each one won.
An agent with its own identity, memory and budget. Or a thousand clones of one.
The 5-question audit
Node › Action › Compound › Rules › Speed
Can there be any viral action? What’s the thing that creates a hook & pulls in another node.. the shared doc, the invite, the PR comment, the referral
The shared doc. The invite. The PR comment. The referral.
The PR comment that loops in a teammate. The task one agent hands to another. The output a human forwards.
The 5-question audit
Node › Action › Compound › Rules › Speed
Does value actually compound with growth? Or does the tenth node just add noise & confusion?
The tenth teammate in the doc makes it richer.
The tenth agent in the channel: richer, or a wall of 👀?
The 5-question audit
Node › Action › Compound › Rules › Speed
Who sets the rules of the loop? Every loop runs through a surface someone controls .. a feed, an app store, an inbox, an API. If a rule changes here, it can kill a loop.
A feed, an app store, an inbox.
An API, an MCP server, a model provider’s policy, a rate limit. Slack’s agents run on Salesforce’s terms. In Buzz you own the room.
The 5-question audit
Node › Action › Compound › Rules › Speed
How fast can the loop spin? Everyone measures conversion rates, but fewer measure time per cycle complete. A loop that turns in hours compounds past a bigger loop that turns in weeks.
Days to weeks. The invite waits for someone to open email.
Minutes. The agent is always on and never waits for Monday.
The 5-question audit
The audit, on one page
Run it on your own product this week to test whether your loops still work when agents join them. One loop at a time.
| # | Question | Score |
|---|---|---|
| 1 | Who’s a potential node?A user, an admin, a team, an agent? | |
| 2 | Can there be any viral action?What’s the thing that creates a hook & pulls in another node.. the shared doc, the invite, the PR comment, the referral | |
| 3 | Does value actually compound with growth?Or does the tenth node just add noise & confusion? | |
| 4 | Who sets the rules of the loop?Every loop runs through a surface someone controls .. a feed, an app store, an inbox, an API. If a rule changes here, it can kill a loop. | |
| 5 | How fast can the loop spin?Everyone measures conversion rates, but fewer measure time per cycle complete. A loop that turns in hours compounds past a bigger loop that turns in weeks. |
Holds: the answer is the same with agents in the loop. Watch: it depends on something you don’t control yet. Breaks: the model has no answer.
What changed underneath
Density › Learning › Cost
Agents break the scarcity assumption. An agent is always on, never churns from boredom, & can be cloned a thousand times without adding any value. This means that density can now be manufactured.
What changed underneath
Density › Learning › Cost
Data network effects, on steroids. Once one agent can tap into what another agent has learned, a single node’s improvement can instantly spread across the network, resulting in every node leveling up.
What changed underneath
Density › Learning › Cost
Every loop cycle now has a price tag. Old loops cost low-to-zero $. A share, an invite, a post is all free. Today every cycle burns tokens, & somebody pays for them.
Who’s paying for the delta between the value created per cycle & the tokens burned per cycle? The agent doesn’t self-select the way a human does. It comments whether or not it has something worth saying.
Live audit
Your loop, scored
Drop one loop from your product in chat. Who acts, what it creates, who it pulls in. We run one through the five questions, live.
| # | Question | Score |
|---|---|---|
| 1 | Who’s a potential node?A user, an admin, a team, an agent? | |
| 2 | Can there be any viral action?What’s the thing that creates a hook & pulls in another node.. the shared doc, the invite, the PR comment, the referral | |
| 3 | Does value actually compound with growth?Or does the tenth node just add noise & confusion? | |
| 4 | Who sets the rules of the loop?Every loop runs through a surface someone controls .. a feed, an app store, an inbox, an API. If a rule changes here, it can kill a loop. | |
| 5 | How fast can the loop spin?Everyone measures conversion rates, but fewer measure time per cycle complete. A loop that turns in hours compounds past a bigger loop that turns in weeks. |
Click a score to cycle Holds → Watch → Breaks. Enter clears the board for the next loop.
Conclusion
The playbook isn’t going anywhere
The questions I’d ask about any network ten years ago are still the right ones today: who’s the node, what pulls in the next one, who sets the rules of the loop.
Density can be manufactured.
Learning can travel between nodes.
Every loop cycle has a cost.
Thank you
The post this came from: Agent Network Effects & Loops