The base model is becoming a commodity. The durable engineering — the moat — is everything wrapped around it: routing, context, cascades, and knowing when a smaller model, or no model at all, does the job.
Six speakers made this argument independently — in the keynotes, the AI Engineering track and the Leadership track. As far as I can tell, none of them compared notes.
¶1On the second morning, Dave Hall and Avni Bhatt were speaking at much the same time, in different rooms, to different audiences — and making the same argument. In the Leadership track, Hall showed a decision tree that reaches for an LLM only as a last resort: “start with an if statement or ten.” Hall ▸ ≈21:00 Next door in AI Engineering, Bhatt explained how a small language model beat the frontier LLM it replaced in a production document pipeline. Bhatt ▸ Neither knew what the other was saying. That sort of thing happened a lot across the two days, and this theme is the clearest case.
¶2George Cameron made the case first, in the opening keynote. His first theme of 2026, straight off the slide: “AI is increasingly not just Models. Model → Harness.” ▶ watch His market data explains why. The gap between frontier models is narrowing — open-weight models trail by three to nine months — and the price of any given level of capability falls 10–100× within months of it first appearing. “It's not two-x cheaper — you can go ten-x cheaper by choosing a cheaper model that has more recently been released.” ▶ spoken, ≈12:22 When capability converges and prices fall like that, choosing a model stops being about loyalty and starts being about routing.
¶3Speaker after speaker made the same point, mostly without meaning to. Navan Tirupathi put it on a slide in so many words: “The model is a commodity; the harness around it is the moat.” Tirupathi ▸ slide Michael Hart gave the discipline its name — “harness engineering… how we build systems around models to turn them into work engines” — and its equation: agent = model + harness. Hart ▸ ≈01:23 Stephen Sennett supplied the economics. “We don't choose. We default” slide — and defaulting to the biggest model means paying frontier prices for Dockerfile work. His working numbers: about 70% of the value comes from mid-range models, and retrying upward on failure beats over-committing on frontier. Sennett ▸ ≈07:14 “Right context beats more model.” slide
¶4Two speakers brought production numbers. Bhatt: “A demo is one document, on a good day, while everyone is watching. Production is thousands of documents, at two in the morning, when nobody is watching.” slide Her architecture is the theme in miniature — SLM by default, escalate the strange 2% to the LLM — and it came back 10× cheaper and 10× faster, with fewer errors. Bhatt ▸ Hall ran the same play with BERT at the heart of a ticket-routing system: everything through a frontier model would cost about US$30,000 a year. His BERT bill: $212. Hall ▸ ≈19:42
¶5Nobody planned this. A keynote with market charts, an agents talk, a cost talk, a bank's production story and a leadership decision tree all arrived at the same conclusion from different directions. That's about as close to consensus as this field gets right now — though not settled entirely, as you'll see below.
“Coding agents — they're not the model, they're the model and the harness. Both are drivers of performance.”
“This sort of discipline has been termed harness engineering… how we build systems around models to turn them into work engines.”
“When I build these things out, I tend to get about 70% of my value out of those mid-range models.”
“By default, go to SLM… if it doesn't comply, escalate to LLM… 98% of the documents go through the SLM.” Result, from the slide: 10× cheaper · 10× faster · errors ~1 in 20.
“If I was running this system with everything through Opus — [$]30,000 US per year. Using BERT: $212.”
“Start on the bottom rung of the ladder. Start with an if statement or ten… classical ML… a text classifier… then reach for an LLM as your last resort. Your budget will thank you.”
“Just switching everything over to the updated model is a really lazy way to be selecting models for your customers — you're giving them a more expensive and slow experience.”
“Not every model needs to do every job… don't lock yourself into one option or one vendor. And don't sleep on open source.”
“A lot of it isn't modeling. A lot of it is engineering. It's infrastructure. It's permissions. It's security.”
“Clients don't want token-consuming monsters in their AI workflow. They need micro language models — very focused reasoning on a very specific, narrow application.”
“People are overbuilding in terms of harnesses and the code around AI… agent harnesses — the simpler, the better. Give the model everything it needs and get out of the way. Way too much hype is around complicated harnesses that are not necessary and actually degrade performance.”
The moat argument and the minimalism argument are both live, sometimes in the same person. What nobody defended was the default: biggest model, everywhere, unexamined. That's the consensus — narrower than the hype, but real.
| Engineers | Route by task, not by habit: default to a mid-tier model, retry upward on failure. Sennett's benchmark: ~70% of the value from mid-range. Trade-off: retries add latency to the failure path. |
|---|---|
| Team leaders | Make “why this model for this task?” a review question that needs an answer. Bhatt's cascade — SLM default, escalate the strange 2% — is the reference pattern. |
| Org leaders | Ask for per-task unit economics, not the vendor lineup. Hall's router: ~US$30k/yr on a frontier model vs $212 on BERT. If nobody can produce that comparison for your workloads, that's the gap. |
| Key takeaway | Stop auto-upgrading every workload to each new frontier model. Sachs: “a really lazy way to be selecting models.” |
Cameron says harnesses are overbuilt; Tirupathi says the harness is the moat. Run the argument for your own stack: which failure mode are you closer to — orchestration you don't need, or defaults you haven't examined?
For now these are placeholders. Once the individual sessions are up on Conffab, each title here — and every quote above — will link straight to the talk: video, transcript and all.