Full Stack

Ali Haghani on why AI agents need a company brain

Ali Haghani· Co-founder & CEO at Circleback
·~15 min·English·Y Combinator
AgentsAI CompanyBusiness Strategy
TL;DR

Circleback's co-founder argues that as AI agents do more work, conversation context becomes scarce infrastructure: capture broadly, share it only with the right people, and reserve consequential judgment for humans.

01Core Mental Model

Without Shared Context, Agents Work in Isolation

Circleback's bet is that a company's real asset is one shared memory of every conversation — the context both people and AI agents need to act well.

otherwise your AI agent is working in isolation. It doesn't know what was said earlier today or what a customer said.

— Ali Haghani, Full Stack (Y Combinator)
Key Insight
Haghani frames memory, not model quality, as the binding constraint on agent usefulness — a bet that whoever owns a company's shared context, not whoever ships the smartest model, captures the value.
02Compounding Uses

One Memory, Many Jobs

Once every meeting is captured and searchable, the same memory quietly becomes an applicant tracker, a CRM timeline, a support archive, and a prep assistant.

I asked CircleBack to prep me for this and it went through all the emails that we've exchanged, past conversations that you and I have had to get a full understanding of like what is this going to be

— Ali Haghani, Full Stack (Y Combinator)
Key Insight
None of these were designed as separate features — they emerge for free once every conversation is queryable. The product's moat is the substrate, so each new use case costs almost nothing to add.
03Org Design

Agents Out of the Silo

Moving agent work out of a private chat and into shared Slack channels turns a founder's personal automation into visible, company-wide infrastructure.

it makes it a company effort and also gives everyone visibility

— Ali Haghani, Full Stack (Y Combinator)
Key Insight
Pushing agent work from a private DM into a public channel is an organizational choice, not a technical one: it trades the founder's convenience for institutional memory and shared oversight of what the agents actually did.
04Coding Workflow

Review Is the New Bottleneck

When agents write code end-to-end, the bottleneck shifts to review — so Circleback spawns different review agents for code quality, architecture, product, and copy.

the bottleneck then becomes code review which we've also done a lot to kind of spawn different types of agents depending on what the PR is

— Ali Haghani, Full Stack (Y Combinator)
Key Insight
The specialist reviewers reveal that code review is a bundle of distinct failure modes: correctness, architecture, product fit, and voice. Agent generation scales only if scrutiny is decomposed with equal care.
05Quality Systems

Evals Beat Whack-a-Mole

Without evals, improving one thing silently breaks another; a scoreboard on every prompt or model change is what ends the endless game of whack-a-mole.

if you don't have good systems in place to make incremental improvements, you're constantly playing this game of whack-a-mole

— Ali Haghani, Full Stack (Y Combinator)
Key Insight
Evals turn prompt-tuning from taste into engineering. Without a scoreboard, every 'improvement' is a coin flip you can't audit — the eval is precisely what lets a small team swap models or rewrite prompts without fear of quiet regressions.
06Guardrails

What Agents Don't Get to Do

Agents can draft, take a first pass, and build to spec — but they may never send, ship copy to production, touch data, or architect; a human still gates anything that reaches the outside world.

Never send an email on its own. A draft is okay.

— Ali Haghani, Full Stack (Y Combinator)
Key Insight
The through-line of his guardrails is reversibility: agents may do anything a human still gates before it reaches the outside world or the architecture. Autonomy is granted in proportion to how cheaply a mistake can be undone.
07The Frontier

Imagination, Not Bandwidth

As the cost of writing software falls, the constraint stops being skills or bandwidth and becomes judgment: what you decide to build and how you compose it.

what matters more is not how fast you ship, but what you decide to build and how you decide to build it

— Ali Haghani, Full Stack (Y Combinator)
Key Insight
When execution is cheap, taste becomes the bottleneck. Haghani's real claim is that advantage migrates from how well you build to what you choose to build — a shift that rewards judgment over raw engineering horsepower.