What I Learned at the Berkeley Agentic AI Summit 2026
9 min read
Lance Ennen
Share:
I spent Saturday at the Agentic AI Summit 2026 at UC Berkeley — the first day of a two-day event (August 1–2) hosted by Berkeley RDI, spread across four stages with names like Plenary, Nexus, Compass, and Atlas, and drawing what organizers projected as more than 5,000 in-person attendees. Two days earlier I'd been at Stanford BASS listening to Visa talk about stablecoins; in a few days I'll be on a plane to Ai4 in Las Vegas. Berkeley is the middle stop, and in many ways the most personal one.
Personal because this summit is about the exact thing my team does every day: running software development through coordinated AI agents. I came in with a specific question — is the way we operate an outlier, or is it where the industry is actually heading? I left with a clear answer, and a longer list of things we need to fix.
My badge from day one of the Agentic AI Summit 2026 at UC Berkeley. Photo: Lance Ennen.
The setting: an agent conference at campus scale
The summit took over a chunk of the Berkeley campus around Zellerbach and the MLK building — the venue map on the big screen marked stages, poster sessions, and sponsor exhibits across Lower Sproul Plaza. The 2025 edition of this event was a single day; this year it needed two days and four parallel stages. That growth curve alone tells you something about where the field is.
The 2026 venue map: four stages — Plenary, Nexus, Compass, Atlas — plus poster sessions and sponsor exhibits across the Berkeley campus. Photo: Lance Ennen.
The opening plenary set the tone with a slide titled "Now We're at a Critical Point," collaging recent public statements from Jensen Huang, Mark Zuckerberg, Demis Hassabis, and Dario Amodei alongside the "Pacing the Frontier" statement — an open letter signed by more than a thousand employees of frontier AI companies calling for coordination as automated AI research accelerates. The framing wasn't hype. It was: the people building this technology are publicly telling you it's an inflection point, and they disagree about what to do next.
A plenary slide juxtaposing public statements from Jensen Huang, Mark Zuckerberg, Demis Hassabis, and Dario Amodei with the "Pacing the Frontier" open statement. Photo: Lance Ennen.
From copilots to coworkers is no longer a metaphor
The thesis I kept hearing, in different vocabulary from different stages, is one I've been living for the past year: the interesting frontier is not a smarter autocomplete. It's systems that plan, execute, review, and communicate across long-running workflows — coworkers, not copilots.
What made Berkeley different from the usual conference version of that claim is that the talks were about mechanics, not vision. Nobody needed convincing that agents can write code. The sessions were about harnesses, memory, permissions, evaluation, and coordination — the unglamorous machinery that determines whether an agent system works twice or just once.
Agent harnesses matter more than model choice
The single most useful technical session I caught was an NVIDIA talk built around a blunt slide: "Agent Harnesses Matter!" — harness design alone moves benchmark scores by double digits, same model.
An NVIDIA session slide on agent harness design, previewing NVIDIA Labs' NeMo Object-Oriented Agents (NOOA) research. Photo: Lance Ennen.
The slide previewed NVIDIA Labs' Object-Oriented Agents (NOOA) research: an agent is a self-modifying Python object — methods are its actions, fields are its state, docstrings are its prompts. State passes by reference instead of serialized text dumps. Deterministic work stays in ordinary Python methods where it's testable. The numbers on the slide — 82.2% on SWE-bench Verified, 86.8% on CyberGym L1, at roughly half the tokens of a comparable harness — matched what NVIDIA later published on its developer blog.
This maps exactly to what we've learned operating our own agents: most of our failures were never model failures. They were harness failures — context assembled badly, state serialized into mush, deterministic steps left to a probabilistic model. The industry catching up to this framing is good news for every team that has been quietly doing harness engineering without a name for it.
The OpenClaw moment
I also caught Peter Steinberger's talk — billed on the agenda as creator of OpenClaw, now at OpenAI. If you saw a lobster emoji fill a giant conference screen, you were in the right room.
The lobster takes the stage: Peter Steinberger, creator of OpenClaw, speaking at the Agentic AI Summit 2026. Photo: Lance Ennen.
This one mattered to me personally, because OpenClaw is the backbone of how my team works. Before OpenClaw, our automation was cron jobs and GitHub Actions — capable, but blind: scripts fired on schedule whether or not the world had changed, and every new workflow meant more YAML. OpenClaw changed the shape of the thing. Our agents now live where the team lives — Slack — with persistent context, and the orchestration layer handles scheduling, memory, and tool access instead of a pile of point-to-point scripts.
Seeing the project's creator on a Berkeley stage, months after the project moved into a foundation with OpenAI's backing, was a marker of how fast this space moved: eight months from a personal side project with a lobster mascot to standard infrastructure discussed at an academic summit.
What running five agents actually looks like
Here's the firsthand context I brought into every hallway conversation, because it's the part people asked about most.
My organization currently operates five AI agents involved in software development. They communicate in Slack — with us and with each other. They participate in our weekly iteration planning meetings and demos, with humans setting direction and reviewing outcomes rather than driving every task. When an agent's work needs course correction, I've literally gotten on Google Meet sessions and talked through feedback — the same way you'd coach a junior engineer, except the notes get folded back into the agent's instructions and memory.
Two honest caveats, because Berkeley's best sessions were honest about the same things:
- This is not "no humans." Humans define goals, review consequential changes, maintain the infrastructure, and own every outcome. The agents expand what a small team can ship; they don't remove the team.
- The failure modes are real. Agents overcommit, lose the thread on long tasks, and occasionally do exactly what you said instead of what you meant. Which is why the summit's obsession with evaluation and observability felt overdue rather than academic.
The validation I came for: our setup — messaging-native agents, recurring planning rituals, human approval gates — matched the patterns the more sophisticated speakers described as where production systems are converging. The challenge I left with: our evaluation story is weaker than our orchestration story, and that gap is now the roadmap.
Security was the quiet headline
The poster sessions delivered my favorite reality check of the day. One poster — "Adversarial News and Lost Profits: Manipulating Headlines in LLM-Driven Algorithmic Trading," from researchers at the University of Liechtenstein and Reykjavik University — showed how imperceptible text manipulations (Unicode homoglyphs, hidden spans invisible to human readers) can flip an LLM trading system's sentiment reading and measurably damage returns, across nine models they evaluated.
A poster session on adversarial headline manipulation against LLM-driven trading systems — invisible-to-humans text changes that flip model sentiment. Photo: Lance Ennen.
Generalize that poster and you get the core security problem of the agentic era: agents read untrusted input and hold real permissions. The OpenClaw project's own security documentation says it plainly: most failures are not exotic exploits — they're "someone messaged the bot and the bot did what they asked." Every serious operator I talked to has internalized some version of this: allowlists, tool policies, sandboxes, approval gates on anything irreversible. If you run agents with production access and haven't thought about prompt injection, you have an incident scheduled; you just don't know the date yet.
What founders and CTOs should take from this
If I compress the day into advice for people running engineering organizations:
- Start with the harness, not the model. Your model choice matters less than how you assemble context, structure state, and separate deterministic work from generative work.
- Put agents where your team already communicates. The Slack-native pattern isn't a gimmick; ambient visibility of agent work is what makes review and trust possible.
- Budget for evaluation like you budget for CI. "It seemed fine in the demo" is not an evaluation strategy. This is my own team's biggest gap and I suspect yours too.
- Treat agent permissions like production credentials. Least privilege, scoped tokens, approval gates, kill switches. The adversarial research isn't hypothetical anymore.
- Keep humans on outcomes. The teams that work don't automate accountability. They automate execution.
The takeaway
A year ago, describing my team — five agents in Slack, weekly IPMs, demos with minimal human involvement — got me polite skepticism. At Berkeley, it got me follow-up questions about our harness design. The conversation has moved from whether to how, and the "how" is maturing fast: harness engineering, permission systems, evaluation, adversarial robustness.
The next question is whether any of this survives contact with the enterprise — which is exactly what Ai4 in Las Vegas is for. That's the next stop.
This article is firsthand reporting from an event I attended, combined with linked primary sources. Photos are my own. See this site's editorial policy for how coverage like this is produced.
Written by
Lance Ennen
CTO & Technical Advisor helping startups and Fortune 100 companies build innovative digital products. Passionate about blockchain, AI, and scalable architecture.
Enjoyed this article? Share it with others