Building AI-native: Inside the stacks powering Cognition, Gamma, and Harvey

Three teams building AI-native products — Cognition, Gamma, and Harvey — discuss the architectural decisions behind their stacks. The conversation covers multi-agent orchestration, MCP in production, autonomous agent design, and the tradeoffs each team has worked through along the way.

May 5, 202628mWatch on YouTube ↗

WHAT IT’S REALLY ABOUT

How Gamma, Cognition, and Harvey architect AI-native products today

All three companies attribute their product breakthroughs to step-function improvements in model capabilities, forcing repeated re-architecture as previously impossible workflows become feasible.
Gamma’s evolution tracks key platform waves—image models, instruction tuning, tool calling/agent orchestration, and MCP connectors—shifting both product capabilities and distribution through embedded “Gamma-as-an-agent” surfaces.
Cognition’s core bet is long-horizon autonomous coding agents, enabled by native code editing, stronger planning, file-system-based “memory,” and rapidly expanding cloud-agent usage.
Harvey highlights multiple inflection points—foundation models, reasoning models, and advanced agentic/coding models—culminating in multi-agent coordination that mirrors hierarchical legal work.
The panel emphasizes durable engineering advantages: robust observability/evals, willingness to delete and rebuild architecture, and infrastructure constraints (security, data boundaries) as prerequisites to organizational—not just individual—productivity gains.

IDEAS WORTH REMEMBERING

5 ideas

Assume you will re-architect repeatedly as models evolve.

Harvey described three major capability inflection points, each requiring a fundamental rebuild; the panel’s shared lesson is that “point-in-time” architectures quickly expire when capabilities jump.

Treat observability and evals as first-class product infrastructure.

Cognition argues that the ability to inspect any agent decision, replay runs, and compare model upgrades is what makes rapid iteration safe; otherwise teams “add a prompt and hope.”

Optimize speed and quality differently by workflow stage.

Gamma is re-orchestrating generation to prioritize “nailed it” quality (users will wait), while keeping editing interactions fast and responsive—making latency/quality dials explicit, sometimes user-controlled.

File-system-native agents reduce the need for heavy custom planning/memory stacks.

Cognition noted that modern models can write plans, navigate repos, and use the file system as durable working memory, moving some teams from bespoke memory/RAG approaches to simpler file-centric designs.

Multi-agent systems should mirror real organizational hierarchy.

Harvey maps partner→senior→junior legal decomposition to coordinated agent hierarchies, which becomes practical with stronger models and orchestration/managed-agent infrastructure.

WORDS WORTH SAVING

5 quotes

Gamma's big bet was that AI-- with, with AI, that we could take away that ninety percent of the futzing that people spend their time on.

— Deanie Fatiha

You build a lot of things that you go delete.

— Walden Yan

You can't make point-in-time decisions and then stick to them.

— Niko Grupin

I've offloaded my entire weekly meal planning- and day-to-day diet to Claude Code. I'm not kidding you.

— Niko Grupin

Individual, uh, productivity gains from AI distributed widely- does not equal organizational productivity gains.

— Niko Grupin

Frontier-model inflection points and product viabilityAgentic workflows and long-horizon autonomyRe-architecture as a recurring operating modeObservability, replay, and evaluation systemsFile systems as agent memory vs. classic RAGMCP/connectors as distribution and acquisitionEnterprise constraints: ethical walls, data boundaries, securitySpeed vs. quality tradeoffs across workflowsMulti-agent coordination and organizational productivity

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.