Claude

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 6, 2026·28m·Watch on YouTube ↗

📖 CHAPTERS

1. 1

   Panel setup: frontier AI architecture questions + founders’ core bets

   Beth Robertson (Anthropic) frames the discussion around the architectural tradeoffs of building frontier AI companies. Each panelist introduces their company and the foundational bet that made the product direction worth pursuing.

2. 2

   Cognition’s early thesis: autonomous coding agents become viable

   Walden Yan explains why Devin’s vision depended on capabilities that didn’t exist at inception. He outlines the model and tooling improvements that unlocked longer-horizon autonomy and rapid growth in cloud agent usage.

3. 3

   Gamma’s capability inflection points: from images + instruction tuning to tool calling

   Deanie Fatiha walks through the ecosystem shifts that made Gamma’s current product possible. Multiple AI waves changed both what Gamma could build and how users experience creation and editing.

4. 4

   MCP and distribution: turning Gamma into an agent inside other tools

   Gamma leaned into the MCP wave to build connectors and embed Gamma in users’ existing workflows. This changed both engagement (frequency) and acquisition (discovery through other platforms).

5. 5

   Cognition’s engineering lessons: from brittle hacks to native code editing + file-system “memory”

   Walden describes the rapid obsolescence of early agent scaffolding and the shift toward leveraging native model capabilities. Practical changes—like relying on file systems for planning and memory—made longer tasks more tractable.

6. 6

   Harvey’s three inflection points: foundation models → reasoning → orchestration offload

   Niko Grupin outlines Harvey’s progression from simple Q&A to complex legal workflows. Each capability jump expanded scope, culminating in models that can take on more planning and coordination responsibilities.

7. 7

   Multi-agent coordination mirrors law-firm hierarchy

   Harvey maps agent systems onto the hierarchical nature of legal work: partners delegate to associates, who delegate further. Improved orchestration infrastructure makes this structure feasible in software.

8. 8

   Hard-earned lesson: be ready to re-architect repeatedly

   Panelists agree that fast model progress forces continuous product and architecture rewrites. Planning must remain flexible, with frequent reprioritization and a willingness to scrap prior assumptions.

9. 9

   Cognition’s caution: invest in observability, evals, and debuggability

   Walden argues the biggest durable investment is not any single prompt or workflow, but the tools that let you understand and improve agent behavior. Without strong instrumentation, teams “build in the dark” and can’t safely upgrade models.

10. 10

    Gamma’s re-architecture: shifting from “wow fast” to “wow nailed it” (and back)

    Deanie explains how user expectations evolved: quality and reasoning matter more during generation, while speed remains essential during iterative editing. Gamma is rebuilding orchestration to flexibly trade off speed vs. quality depending on workflow.

11. 11

    Today’s bets: cloud agents, organizational productivity, and collaboration interfaces

    Walden predicts a shift to AI-driven software orgs where agents run projects end-to-end and pull humans in only when needed. Niko emphasizes that individual AI productivity doesn’t automatically translate to organizational gains, making collaboration and governance central.

12. 12

    Harvey’s near-term focus: infrastructure, data boundaries, and ethical walls

    Rather than only pushing frontier intelligence, Harvey prioritizes constraints required by legal customers. Sensitive data handling, isolation, and governance must be foundational before delegating more autonomy to agents.

13. 13

    Gamma’s forward bets: “taste” and designing for an agent-mediated world

    Deanie argues product differentiation will increasingly come from taste and design range, not just generation. Gamma is also preparing for a world where agents use products as primary users, reshaping UX for human-agent collaboration.

14. 14

    Lightning round: personal unlocks, predictions, and founder advice

    The panel closes with fast takes on practical AI wins, near-term predictions, and strategic defaults. Themes include tool consolidation, job shifts, ambient intelligence, and shipping fast while hiring great people.