Founder Chats - Daulet Amirkhanov

Episode Summary

Executive Summary: The episode traces Dalit Almarkinoff’s path from high-throughput reliability work at Meta to graph-memory and MCP infrastructure at Cogni, and now founding engineering at Beat AI. He argues that agentic systems are becoming capable enough for compliance audits, but robust sandboxing, strong evals, and eventually agent identity/capability registries will be essential for safe, scalable deployment.

Main Topics: Career progression from Meta to Cogni to Beat AI (Priority: 5/5): Dalit explains how infrastructure work at Meta shaped his engineering fundamentals, how Cogni exposed him to AI agent infrastructure, and how Beat AI combines both backgrounds in compliance automation. Meta-scale reliability and shipping discipline (Priority: 5/5): He describes working in Meta ads logging/reliability at massive scale and says the main lesson was breaking problems into small, shippable pieces to validate hypotheses quickly. Agent infrastructure as the next distributed systems frontier (Priority: 4/5): Dalit argues that agent systems are not just hype; they introduce real distributed-systems concerns such as retries, competing requests, and shared state across processes. Cogni’s decoupled shared-memory architecture (Priority: 5/5): He details decoupling MCP from the core memory layer so multiple agent frameworks can share a unified knowledge graph through a standalone process, enabling cross-framework interoperability. Why compliance is a strong wedge for autonomous agents (Priority: 4/5): At Beat AI, Dalit believes compliance audits are a good fit because current agent capabilities are already strong enough to handle structured, high-signal work with clear constraints. Safety, evals, and sandboxing (Priority: 5/5): He emphasizes sandboxing as the only dependable safety boundary today and identifies evals as a continuing challenge because real customer workflows are hard to fully reproduce offline. Future need for agent identity and accountability (Priority: 4/5): Dalit predicts that as agents take on more responsibility, the industry will need identity and capability registries to attribute actions and regulate autonomous behavior.

Key Arguments: Meta’s high-scale ads infrastructure taught Dalit to ship small, validate fast, and think in terms of reliability under extreme load. Agentic systems create familiar distributed-systems problems, so infra engineers with reliability backgrounds have an advantage in building them. Decoupling MCP from a single runtime lets multiple AI frameworks share one memory layer, which is crucial for multi-agent and multi-framework workflows. Compliance is a practical early market for agents because the technology is already advanced enough to automate structured audit-related work well. Sandboxing is currently the only safety mechanism he fully trusts for autonomous agents, especially given supply-chain and data-risk concerns. Evals remain one of the hardest problems because synthetic or local test cases rarely capture every client-specific production nuance. Future agent infrastructure will likely require identity, provenance, and responsibility tracking so actions can be audited and attributed.

Data Points: Time at Meta: a bit over 3 years - Dalit says he worked at Meta for just over three years before moving on. Scale of logs at Meta: millions of logs daily - He says the Shops Ads logging system handled millions of logs per day, only a small fraction of overall ads traffic. Number of frameworks supported at Cogni: around 5 popular agentic frameworks - He notes Cogni was already supporting multiple frameworks before the decoupling work. Working timeline at Beat AI: January (joining as founding engineer) - He says he joined Beat AI in January as a founding engineer. Discount code: 10% off - Promotional mention for Plaud AI using code CodeSTORY.

Pivotal Quotes: "I think right now, sandboxing is the only real safety boundary for autonomous agents." — Dalit Almarkinoff: He gives his contrarian view on agent safety and argues that prompts alone are insufficient. "That mindset of chunking your problem into smallest digestible pieces that you can ship as fast and check your hypothesis as fast with, that's stuck with me ever since." — Dalit Almarkinoff: He reflects on the biggest engineering lesson learned from Meta-scale systems. "I think there's going to be a necessity maybe to have some sort of a identity and capability registry for individual agents." — Dalit Almarkinoff: He predicts future infrastructure will need agent identity and accountability mechanisms.

Implications: For builders, the episode suggests agent infra is moving from experimentation to production discipline: strong safety, rigorous evals, and shared-memory interoperability matter now. For the industry, compliance may be one of the first durable agent markets, while attribution and regulation will likely become core infrastructure needs.