The modern guide for implementing AI agents

This post originally appeared on the Confluence.VC weekly blog.

Where is the world heading? Over the past few months, I’ve written a few pieces on my site Confluence.VC that keep coming back to the same conclusion:

I think the trendline is pretty clear:

• Managing people → orchestrating agents
• Adding headcount → adding loops
• Shipping more features → faster decision-making velocity

Here’s how I’m thinking about it.

What a scaled-out product team actually looks like

Early-stage companies are resource constrained where it hurts most: product. You need the analysis of a big-company growth team and the pace of a three-person startup.

If you’re lucky enough to work at or with a world-class product team, you’ll notice a few things:

• Every decision is centered around the question of whether or not this will improve the end user experience.
• Answering that question comes from making sense of an abundance of data.
• Gathering, analyzing, and implementing that data compounds on itself daily.
• Speed is rewarded.

All of that seems pretty straightforward, but I’ve learned that most companies don’t have systems in place to make this plan a reality.

That’s why I love companies like Amplitude.

They offer AI tools and agents that behave like interns for every job-to-be-done across your product loop:

• Their Dashboard Agent keeps a constant watch on the metrics that matter and surfaces opportunities the moment they appear.
• Their Session Replay Agent automatically detects where users struggle, surfacing actionable insights that help teams quickly validate and fix recurring problems.
• With AI Feedback pull in comments from everywhere your customers are talking (surveys, support tickets, calls, reviews, Reddit, Discord, X, G2), and let AI do the heavy lifting to show you what really matters.
• Their MCP Server pipes Amplitude’s behavioral context into where you already think Claude, Cursor, Github, and more—so you don’t context switch to see what’s happening.

What used to take weeks now takes hours.

When cycles collapse like that, early movers are able to create compounding flywheels that give them more leverage by the day.

Old flywheels vs. new flywheels

I believe that every company has a flywheel. The best companies have flywheels that compound faster than their peers.

Most startups still operate the classic loop:

This works until it doesn’t.

• Signal gets lost in the hand-offs.
• Valuable time is wasted waiting around for previous step to be completed.
• Things slow down every time you need human involvement.

The new loop looks a bit different:

Instead of needing humans at every part of the process, agents run each phase in parallel threads. Instead of incremental improvements, you create continual improvements.

The visual above from Build to Launch shows the real advantage of these flywheels, and that is the continuous feedback and self-optimization that is created when done correctly.

Traditional flywheels compound linearly; agentic flywheels compound exponentially.

Department compression (2.0 → 3.0)

As I have written about in the past, I believe that all departments are going through an evolution. The next phase of that evolution is currently taking place.

• 1.0: Humans solving for different problems at every step of the way. Slow feedback cycles are a byproduct of all of the work needed to manage people + systems.
• 2.0: A small collection of decision makers create more output by using software to automate mundane tasks, unlocking more time towards high-leverage tasks. Feedback cycles improve as systems become easier to manage without more throughout the organization.
• 3.0: Decision maker(s) + agents. The jobs of analysts and many individual contributors disappear, but their work does not; they operate through agents, setting goals, constraints, and taste, then reviewing higher-quality proposals instead of pulling raw data.

Proof (how product teams are actually building out agent workflows)

Replit is built on speed; product learning has to match.

Replit has a really large global network of users. A product like AI feedback makes it really easy to bring all of their data into one place so that every member of their company has access to it.

Replit is an engineering-first company. They have ~40 engineers and only 2 product managers.

Given this dispersion of talent, a previous challenge was getting all of the engineers on the same page. That has now been solved, and the AI feedback product allows each engineer to understand the “why” behind each product decision they make. At scale, this allows Replit to move incredibly fast by removing decision-making bottlenecks in their previous workflows.

Here’s the workflow their team used:

1. Set up and integrate with existing sources like Zendesk, App Store reviews, Reddit, Twitter, etc.
2. Start collecting feedback from these different sources dynamically, grouping that feedback, and creating a prioritized stack of things to fix for the engineering team

What used to be a massive timesuck for the engineering team (figuring out what to build that users are requesting) now takes almost no time.

How to roll this out for yourself

1. Choose your agent infrastructure. I’d recommend Amplitude for this step. They also have a startup program where you get free product and web insights for a year (linked).
2. Deploy one agent. If you’re early and don’t have clean experimentation habits yet, start with Dashboard Agent. Let it watch your core metrics and tell you when something moves.
3. Give your agent context. This is the part people skip. Tell it what good means for your company. “We care about users who complete 3 projects,” “We care about orgs that add 5 seats in 14 days,” etc. The more you talk to it, the more it reflects your strategy instead of generic “improve conversion” noise.
4. Pipe it into where you work. Use Amplitude’s MCP so your PMs see Amplitude insights in Claude / Cursor / Github. The less tab-switching you require, the more the team will actually use it.
5. Make it a ritual. Add “agent insights” to your weekly product review. “What did the agent see this week that we didn’t?” That’s how you make the system compounding.
6. Clone it. Once one agentic loop works (say, onboarding → week-1 activation), copy the pattern for feature adoption, then for monetization.

Wrapping it all up

If there’s one piece of startup advice that I subscribed to myself, it is this:

If you listen to some of the best founders, they’ll tell you the same thing; they’re obsessed with speed.

Speed can only happen with direction, and direction can only happen with compounding loops like the ones I have described throughout this piece.

Those compounding loops used to require massive amounts of headcount. Now they do not, and any company can build a more scalable system using agents.

I have played around with a lot of agent infrastructure for myself, and I have been recommending more and more founders to build with Amplitude.

They have a free plan available where you pay $0 and get 1.2 billion events. (I don’t think you’ll find a better offer for something like this.)

Take advantage of that offer here and start building your agentic workflows for free.