| // PRE-CONSENSUS | NO. 04 · NOV 1 | ||
The Frontier Syndicate |  | ||
// 01 FROM KRISTOPHER
From vibe coding to agentic engineering
 | If AI can write code for almost nothing, what's left for engineers to do? I put that question to three technical leaders who build with agents every day, in a panel at the SVB Experience Center in San Francisco, hosted by Frontier Syndicate, Silicon Valley Bank and Constructive. We got into where vibe coding ends and agentic engineering begins, and what the job looks like now. |
I sat down with:
 Sanjeev Dhanda Sr. Staff Software Engineer, Google DeepMind |  Abhi Aiyer Founder & CTO, Mastra |  Dan Lynch Founder & CEO, Constructive |
My top three takeaways:
1. Writing code has gotten cheap. Trusting it has not. Code generation has become a commodity. The scarce part now is knowing you can depend on what was written. That shifts the bottleneck from writing code to trusting it.
2. Vibe code the edges, engineer the core. Sanjeev called vibe coding "the microwave of cooking": fast and convenient, but no six-course meal. Dan pointed to the bank around us. You can vibe code the lobby, but "you would probably not want your vault to be vibe coded."
3. The team is now one engineer and their agents. Abhi described Mastra moving past pods and squads, with each engineer running an army of agents. The bar for every engineer goes up, because "you can't blame Claude if you broke something."
// 02 INVITES
Upcoming
If a model can generate a world, an agent can practice in it. In late 2025, DeepMind's SIMA 2 agent kept improving inside worlds generated by Genie 3, with no new human demonstrations. Whether that scales comes down to two open questions: can a generated world hold together for more than a few minutes, and do skills learned inside it carry over to the real world?
Bonnie Li is a Research Scientist at OpenAI. At Google DeepMind she worked on Gemini, the Genie world models and SIMA 2, and she began her research career at Mila with Yoshua Bengio. She'll walk through what it takes for a world model to be a place agents actually learn, and the failure modes that matter, like agents that learn to exploit the simulator instead of the skill. Dinner and paper discussion at SignalFire; pre-reads go to confirmed guests.
Ankur Goyal, co-founder and CEO of Braintrust, sits down with Corinne Riley, Partner at Greylock, on why the focus is moving from standalone evals to agent observability, rebuilding databases for agents, and choosing models on data when there are more of them than ever. Part of Baseten's Heavyweights series, at Baseten, 560 Davis St, San Francisco. The event is full; RSVP to join the waitlist.
At The Ritz-Carlton, San Francisco.
Three days of AI for science at The Ritz-Carlton, San Francisco, hosted by the Chen Institute and Bay AI Circle, with 100+ scientists speaking. Thursday is the day to go: Yejin Choi on reasoning and discovery, and Terence Tao on Math 2.0, followed by a fireside chat with Dashun Wang. 1-day passes available. Full agenda.
The robotics research crowd is back at Mission Robotics for the Embodied AI track, co-hosted with New Theory. Most robot learning starts from human demonstrations. Christian Shewmake, Founder of New Theory, takes robot imitation learning from the builder's seat. Alex Proshkin of UC Berkeley (Pieter Abbeel's lab) presents Tactile Curiosity: a robot that explores driven only by curiosity about what it feels, with no task rewards or demonstrations, then uses that touch data to improve VLA models pre-trained without any touch.
VetsinTech Fall Gala
VetsinTech's annual Veterans Day gala at the Marines' Memorial Club in San Francisco honors Victor Glover, a retired Navy Captain and test pilot who piloted Crew-1 and NASA's Artemis II. As VetsinTech's SF Bay Area Chapter Lead, I'd love to see this community in the room.
Reception and silent auction at 6, dinner and program at 7, Cocktails with Veterans from 9 to 11. Individual tickets are $500, and student veterans attend free. Bronze ($10,000) and Silver ($25,000) table sponsorships close Fri Nov 6 at 11:59 PM PT. The night supports VetsinTech's work connecting veterans and military spouses with education, jobs and entrepreneurship in tech. Event details.
Coming up
// 03 FROM THE RESEARCH ROOMS
From the Research Rooms: FRC #29, Grading AI Negotiators With No Answer Key
Thu Oct 22 at Array Ventures in San Francisco, on Learning Human Preferences. Dinner, a research talk, then a long discussion. Thank you to Array Ventures for hosting and GMI Cloud for sponsoring.
FRC #29 at Array Ventures, San Francisco. Photo: Fiona Duong
TERMS-Bench: Evaluating LLMs in Semi-verifiable Domains
 | Erica Zhang is a Ph.D. candidate in Stanford's Department of Management Science & Engineering, co-advised by Jose Blanchet and Mert Pilanci, and a 2026-27 Jump Trading Fellow in AI/ML. |
Math and code have answer keys, so you can grade a model automatically. A negotiation doesn't. TERMS-Bench, built at Stanford with Susan Athey, James Zou, Jose Blanchet and collaborators, makes the environment itself the grader.
The setup: an agent bargains over one price against a simulated counterpart whose reservation price, urgency and stance are hidden from the agent but known to the evaluator. Every model plays the same 1,800 seeded episodes across six counterpart types, from candid to adversarial.
Deal rate hides what matters: apart from GPT-4o-mini, every model closed at least 93% of the deals that were possible. Yet the share of available value captured varied 3.7-fold across the 13 models, and the simplest fixed-concession rule beat GPT-4o-mini.
The surprise: every model kept less value when the other side's words were informative. Warm cues led to over-conceding, and pressure cues triggered brittle play.
More information didn't fix it. Handing a model the exact Bayesian read on its counterpart closed only 1% of the gap to the optimal policy for Claude Opus 4.7.
In dollars: across 192 simulated purchasing scenarios, Claude Opus 4.6 earned $68,592 against $48,776 for the best fixed-concession rule on the same scenarios.
Why it matters: agents that close every deal can still be bad negotiators. Before companies let agents bargain with real money, they'll need evaluations that show where value leaks, not just a deal rate.
// 04 FRONTIER FORECAST
Frontier Forecast
Real signals, played forward into one plausible future. Each ends with what to build, what to invest in, and what to watch.
By Kristopher Floyd
TOPIC // AGENTS THAT NEGOTIATE
SCENARIO: Both Sides of the Table // 2030
It's 2030. A mid-size manufacturer needs freight for 400 lanes next quarter. Its buying agent opens the event at 6 AM. By 8, carrier agents have bid, countered and settled most of it, and no person has typed a word. People show up for the exceptions: a new carrier, a lane where the agent walked away, a supplier whose agent kept conceding when the buyer sounded friendly. The procurement lead's dashboard doesn't track deal rate anymore, because everything closes. It tracks value left on the table, per agent and per counterpart type, and which model to retire next.
Signals
Keelvar says AI agents now run 90% of sourcing events on its platform, up from about 71% a year earlier. Supplier-side agents bidding into buyer-side events reached 1,420 events in July 2026, about 8% of the total, and Keelvar predicts more than half of its events will be machine-to-machine by the end of 2027 (company figures, Procurement Magazine, Oct 28). Why it matters: agent-to-agent negotiation is already live in freight, parcel and packaging, where pricing is formulaic.
TERMS-Bench, presented at FRC #29: across 13 frontier models, deal rates were close to saturated while the value captured varied 3.7-fold, and warm cues from the counterpart made agents concede more. Why it matters: the failure that costs money is invisible to the metric most teams track.
Trend
Negotiation is becoming software on both sides of the table, and the scoreboard is moving from whether a deal closed to how much value it kept.
From: human buyers, quarterly RFPs, deal rate and savings as the scoreboard.
To: agent-to-agent events that settle in hours, with mechanism design and evaluation deciding who wins.
If this future happens...
🛠️ Build: evaluation suites that score agents on value captured and walk-away discipline, simulators that model real counterpart types, replay tools that show why an agent conceded.
💰 Invest: teams with proprietary negotiation outcome data, marketplace and mechanism-design infrastructure for agent-to-agent trade, guardrail layers that enforce price and term limits.
👀 Watch: the share of machine-to-machine events on sourcing platforms, whether model providers start publishing negotiation evals, and the first public dispute over a deal two agents struck.
// 05 OPEN CALLS
Open calls
Speakers: got a paper worth arguing about? Bring it to the room.
Sponsors: back the sessions where the Bay's researchers, founders and funds meet.
Compute: building something that needs B200 or B300 nodes, colo or power? Tell us what you need.
// STAY IN THE ROOM
Know a builder who should be in these rooms? Forward this.