How Unify cut its AI agent costs 95% in two weeks — Unify CTO on Max Agency podcast
LangChain · x · 2026-09-03
LangChain's Max Agency podcast hosted Unify co-founder and CTO Connor Heggie on the engineering behind GTM agents. Unify's philosophy: not replacing sales reps, but giving every rep 'an engineer in their back pocket.'
Key takeaways:
- Cut agent costs 90-95% in the two weeks before launch
- OpenAI's prompt cache hides a 15-requests-per-second ceiling; optimizing cache hit rates is the key to savings
- Your LLM judge must run on a different model family than the agent it grades — same family causes mode collapse, essentially groupthink for agents
- The Speed Audit: auditing one at a time beats a table of 1,000
- Subagents are just function calls; fork-vs-child subagent trade-offs
- Replaced full VMs with Monty, a Python REPL, for durable cloud agents; gave models pandas-like powers over a live table; locked memory to keys instead of letting the agent freestyle
- Heggie's background: a 15-person self-driving startup and Scale AI's mapping team shaped his eval methodology
A one-hour deep engineering retrospective with high practical value.
Related event: Unify CTO Shares Agent Engineering Rules, Cutting 95% Cost in Two Weeks(2 posts)→
More from coding & agent
- MongoDB Atlas now powers the virtual file system behind LangChain's Deep Agents — BraceSproul · 2026-09-04
- Omnigent Ships LLM Cost Controls: Smart Routing Plus Progressive Budgets to Stop Runaway Bills — TerryTangYuan · 2026-09-04
- E2B sandbox runs RL rollouts up to 3x faster, cutting idle GPU time and training cost — badphilosopher · 2026-09-04
- Claude Code adds /limit-reset, letting users manually reset the 5-hour usage cap once per week — dotey · 2026-09-04
- LM Studio Launches Bionic, a Fully Local Agent for Work and Code — mattturck · 2026-09-04
- Perplexity launches fully local Portable Computer on Linux RTX GPUs and DGX Spark — AravSrinivas · 2026-09-04