Build Daily
Tinley Park · August 23, 2026How Unify cut its AI agent costs 95% in two weeks
Connor Heggie, co-founder and CTO at Unify, explains how the company cut its AI agent costs by 95% in two weeks before launch by moving from expensive sub-agents to a smarter main agent with prompt caching.
The outline
Tap a timestamp to jump there in the video.
Intro & cost optimization
Connor Heggie introduces Unify's outbound agent platform and the 90-95% cost reduction achieved in two weeks before launch.
Prompt caching & LLM judges
Heggie discusses why prompt caching is critical for cost control and why using an LLM as a judge risks mode collapse and groupthink among agents.
Company knowledge in memory
He argues that company-specific knowledge should live in memory rather than skill files, since most non-engineers won't write them.
Unify's go-to-market thesis
Heggie explains Unify's core idea: go-to-market is a search problem over unstructured data to find companies with problems your product uniquely solves.
Early web-agent research system
The first version used scaled web agents in reflect/tool-call loops to answer structured questions like whether a company performs KYC.
Planning step with reasoning models
When O1 arrived, Unify added a planning step that generated strategies and pitfalls, massively improving research quality over the long tail of their TAM.
Scalability & async workflow UX
The original product ran agents asynchronously on tables of thousands or millions of records via a drag-and-drop workflow builder, optimized for low cost at high scale.
New chat product for sales reps
Unify launched a chat-based agent that gives individual sales reps an engineer in their back pocket to hit APIs and build tables dynamically.
Agent harness vs. coding agents
The harness is similar to coding-agent harnesses but outputs database records and emails rather than file-system edits, still achieved by writing code in a sandbox.
Cloud durability & TypeScript criteria
Because Unify runs in the cloud for non-technical sales reps, the harness must be durable, background-capable, and write TypeScript code.
A solo builder can ship the same outcome with fewer agents and tighter prompts, not just more compute. Cutting 95% of cost in two weeks by consolidating sub-agents into a smarter main agent shows that efficiency is a design choice, not a post-launch fix.