Day 3's throughline was harness engineering — the MC's framing that "the real problem with building with agents is not the model but everything around it." The morning stacked the State of AI Engineering survey, Stanford's Homa low-latency protocol, DSPy's separation of task from model, a Mike Krieger fireside on building with Fable, and Neo4j's semantic layer; the afternoon ran the Great Loops Debate into talks from Theo, Garry Tan, and Howie Liu, closing with the Startup Battlefield won by Comment.io.
MC Ralph Shabri (Replit) framed Day 3 as "harness engineering," summarizing the morning as "the real problem with building with agents is not the model but it's about everything around it." Amplify Partners' Bar opened with the 2026 State of AI Engineering survey — 1,048 respondents, run for the first time with Notion and Vercel. Her headline shifts: image-generation adoption doubled (18%→36%), audio has the highest intent to adopt, 95% of teams now use agents with 89% of those agents able to write data (so write-enabled agent use more than tripled), and "cost is now a first class engineering constraint."
John Ousterhout (professor emeritus, Stanford) made the case that latency is starting to matter more: inference and agentic workloads increasingly exchange small metadata and coordination messages, where legacy TCP and RDMA suffer high tail latency from incast congestion. He introduced Homa, a clean-slate data-center transport that is message-based (not a byte stream), controls congestion from the receiver, and uses switch priority queues to favor short messages — reducing tail latency by roughly 13x in his benchmark. Now semi-retired to work on Homa full-time, he invited the audience to try the open-source Linux kernel module.
Maxime Rivest & Isaac Miller (DSPy) argued for "the unreasonable effectiveness of separating the tasks from the model." Treat a repeated AI task like a function with a fixed input/output contract, then swap prompts, agents, tools, and loops behind it. Fully specifying a task needs three things — instructions (what should happen), code (what must happen), and evals (what good looks like) — after which you can automatically optimize. They cited enterprise wins like Shopify going 550x cheaper by moving to a smaller model behind the same interface, and previewed DSPy 4 features (dspy.flex and qualitative learning).
In a fireside, Mike Krieger (Instagram co-founder; MTS at Anthropic) described how building with Mythos/Fable pushed him from step-by-step task delegation to "express the end state and then have it go and cook on it" — including porting a Python codebase to TypeScript over a weekend. He described Anthropic's internal "tag" workflow as multiplayer, async, and proactive, being "bottlenecked on human ability to even fully conceptualize what we're doing," and stayed bullish on startups because "writing code was never the limiting part." Emil Eifrem (CEO, Neo4j) closed the morning with a blueprint for scaling enterprise agents: "thin agents on a smarter shared substrate," built from a business-facing ontology, a technical ontology of data sources, and runtime execution traces that let agents learn.
The afternoon opened with the Great Loops Debate (Oxford format, hosted by Ally How of Keycard), teams split on whether there is a delta between the hype behind loops and what works in practice. Ian Livingstone (Keycard) and Jeffrey Huntley (creator of the Ralph loop) argued loops are inevitable and are the core unit of engineering with the right discipline, infra, and tests; Dex Horthy (HumanLayer) and Greg (Sentry) argued the hype is outrunning the discipline and that a software factory can run mechanical, spec-gated, test-covered slices but "cannot autonomously decide whether it built the right thing."
Highlights included Huntley's insistence that engineers must "engineer away those failure domains" because "the models are drunk" and can't be trusted, Livingstone's point that goal-seeking RL models are inherently hard to keep aligned, and Greg's repeated worry about the economic viability of stacking loops on loops. The mind-change vote at the end was too close to call — "the delights are so bright" — echoing last year's MCP debate where the debaters half-switched sides.
Theo (t3.gg) urged builders to "go bigger," reading models as eras — Sonnet 3.5 as the reliable-tool-call era, Opus 4.5 as long-running tasks, and Mythos as orchestration — and argued engineers are stuck in a "skeuomorphic phase," over-identifying with tools and languages. His tiers have all shifted down a level: yesterday's startup is today's side project, and some products are now "just a markdown file running on a cron." His closer: "If your idea doesn't feel stupid, it's because your idea is not big enough."
Garry Tan (CEO, Y Combinator) claimed a ~400x personal output jump and insisted "it's not the model. The 2x people and the 100x people are using the exact same claude." He mapped agent tooling onto an org chart — a skill file is an employee, a resolver table is an org chart, filing rules are process, evals are performance reviews — and pitched the "company brain" (his open-source GBrain) as the memory layer, with the mantra "never do one-off work... skillify it." Howie Liu (CEO, Airtable/Hyper Agent) followed with "hiring employable agents," tracing the form-factor progression from completions to chatbots to agents to always-on "claws" to orchestrators, illustrated by a landscaping business running end-to-end on a fleet of agents that learn from every interaction.
Hyper Agent's Startup Battlefield narrowed 500 founders to 20 competitors and three finalists, judged by Theo and Joshua (Hunen), with $100K in prizes. Kamad pitched an agentic execution layer for physical commodities trade (governed by its patent-filed CFC state machine); Comment.io pitched a multiplayer markdown editor built for both people and agents; and Built by Foundry pitched turning content creators into founders by having agents find a painkiller their audience will pay for — a claim Theo pushed back on from lived experience.
In the results, Kamad took second runner-up ($20K), Built by Foundry took runner-up ($30K), and Comment.io won the $50K grand prize. MC Ralph Shabri then closed AI Engineer World's Fair 2026: the most ambitious edition yet at 4 days, 7,000 attendees, and 40 tracks, with thanks to sponsors led by Microsoft and to the backstage crew — before teasing the next AIE (finance-focused, in New York).
It's not the model. The 2x people and the 100x people are using the exact same claude. Same weights, same context window, same API.
If your idea doesn't feel stupid, it's because your idea is not big enough.
Model quality is rented, but if you build your brain, you own that brain.
The real problem with building with agents is not the model but it's about everything around it.
| Time | Topic |
|---|---|
| 00:16 | State of AI Engineering survey — Bar, Amplify Partners |
| 00:36 | John Ousterhout (Stanford) on the Homa low-latency protocol |
| 00:54 | DSPy — separating task from model (Maxime Rivest & Isaac Miller) |
| 01:12 | Mike Krieger (Anthropic) fireside on building with Fable |
| 01:40 | Emil Eifrem (Neo4j) on ontology-based semantic layers |
| 03:40 | The Great Loops Debate (Oxford format) |
| 07:42 | Theo (t3.gg) — go bigger |
| 07:58 | Garry Tan (YC) — build the AI-native company |
| 08:19 | Howie Liu (Airtable) — hiring employable agents |
| 08:37 | Startup Battlefield finale & conference close |