AI Funders, Builders, and Researchers in the Summer of '26
I was laid off at the end of January 2026. Perfect timing.
According to Nathaniel Whittenmore, author of the AI Daily Brief, this was at the start of ‘the month that AI woke up’. According to NJW, all those execs and developers who had been tinkering with Claude Opus 4.5 over the Christmas break came back to work and started to flex their new-found superpowers in new and exciting ways. Things were beginning to happen. Peter Steinberger created -OpenClaude-, -Moltbot-, OpenClaw: an agent that just ran on any machine you installed it on. It could do almost anything on that machine - including delete your email and max out your credit card, if you weren’t careful.
To quote a passage from ‘The Stars My Destination’ - the sci-fi masterpiece where my name comes from:
“This was a time of high adventure, rich living, and hard dying … but nobody thought so”
I was starting to worry that I’d miss the party. I dug into my own open source ideas to keep busy and built a agentic curation system called ‘Skillful Alhazen’. I reached out and an old colleague, Paul Groth, now a fully-fledged professor at the University of Amsterdam, showed an interest in the work and invited me to join his group as a guest researcher. As an aside, he suggested that we submit a demo paper to the upcoming inaugural ACM Conference of Agents and Agentic Systems and I thought ‘why not?’. We scrambled a bit, prepared some demos, wrote the paper, and submitted. We were delighted when we learned a few weeks later that it had been accepted.
So then, at the end of May, whilst jobhunting for AI-Agent-related biotech positions in Silicon Valley (which is a subject deserving of its own bleak dystopian essay), I attended CAIS 2026. As an ex-academic researcher in my mid 50s, it felt wonderful to return to an academic meeting, not to mention a little daunting and exhilerating given that studying ‘Agents and Agentic Systems’ academically whilst the rest of the world is using this technology at a breakneck pace, spending billions in the process and remaking society as it goes. And yet, here we were, a room full of scientists, trying to understand the space and how it all worked.
I went to CAIS with a mindset focussed at the ‘top of the AI stack’ (think of layers consisting of model pretraining / model postraining / applications etc.) with the clear intention of networking and looking for a job. This write up is colored by that focus, but I will try to cover aspects of the meeting that stood out as worthy or interesting.
Since then, I have found a job, but felt that the CAIS26 meeting was a wonderful opportunity to explore the unfolding space of this thrilling discipline. Thanks to Prof Groth for giving me the chance to collaborate and attend!
Talks & sessions — my observations
Andy Konwinski keynote — “Keeping AI Open”
The first keynote was Andy Konwinski (A Databricks Founder) making a comparison to the early days of the Internet and how people had to fight at the time to keep it open. His argument was that the open substrate won because anyone could build apps on it. To me, it seems like the equivalent idea is that AI’s real opportunity is how it applies to different knowledge domains, and that the value accrues to knowledge-based, domain-specific systems, not business apps. Andy challenged us all to place ourselves in the AI Stack and to connect others at the meeting who were located in other parts of the stack. This was a call to action - to fight to try to keep the technology open. I found (and still find) it hard to see how that could be remotely possible given the sheer amount of money involved - and the power of the frontier labs to drive the conversation.
ACDL — A Language for Describing Agentic LLM Contexts
(link) This work dealt with reverse-engineering context — taking the tangle of instructions, observations, and interaction history an agent actually receives and giving it an explicit, shared, diff-able structure. It targets the gap that makes agentic systems hard to communicate and reproduce: there’s no standard way to say what the model is actually being shown and how that changes over time.
TraceFix — Verification-First Multi-Agent Coordination
(Best Paper, link) I spoke with Jorge Ortiz (Rutgers ECE). He frames his smart-cities agent program as requiring agent claims to be substantiated by evidence — which was precisely consistent with my work on scientific curation. TraceFix (counterexample-driven repair of coordination protocols) is his formal-methods instantiation of that framing at the coordination layer.
Terminal-Bench 2.0 / RLEval workshop — the one that stuck with me
I attended the workshop on Terminal Bench, which was fascinating. I asked the question: how do you build benchmarks for knowledge domains not inhabited by the “red-pilled” early-adopter crowd? TB2 is crowd-sourced, but the crowd that writes terminal tasks is overwhelmingly terminal-native engineers — SWE/sysadmin/ML/security dominate, while the long tail of expert work (clinical, legal, lab biology, ecology, accounting, civil engineering, the humanities) is barely present. The benchmark measures what its contributors are, not the full space of valuable work.
Personally, I think this is also a big part of why non-AI people resent AI: the systems are evaluated by, optimized for, and validated against a narrow self-selected technical in-group. To everyone outside it, AI shows up as something done to their domain — never built with them, on criteria they’d recognize as competent. The coverage gap and the legitimacy gap are the same gap. I chatted to Alex Shaw (Laude Institute; co-creator of Terminal-Bench + Harbor) about this and we seemed to agree that it might be possible to incentivize experts to contribute by presenting their participation as a way to steer frontier models into your domain. A benchmark is a gradient signal the labs optimize against — so the incentive isn’t altruism or pay-per-task, it’s agenda-setting power over what the models become good at. Experts who writes the evals sets the frontier, the game, and the stakes.
Alex Dimakis @ RLEval — Open Thoughts & Terminal-Bench Science
Dimakis recommended Terminal-Bench Science as a contribution venue for benchmarks.
Why Johnny Can’t Use Agents
(link) One of the only talks that studied how normal people actually use agents rather than how agents score on benchmarks — the human-factors counterweight to everything else at the conference, and the thing the supply-side reliability papers implicitly ignore.
Malice in Agentland
(link) Leo Boisvert (ServiceNow Research / Mila) presented a scenario that is genuinely alarming: TM3 environment poisoning is persistent and distills into the student model’s policy, and every tested guardrail model plus a weight-based defense failed to detect or prevent the backdoor. That result should change how people think about deploying distillation-based pipelines.
Tressoir — Interpretable Blueprints
I read this to decide how Alhazen could adopt the Interpretable Blueprint structure. The finding that landed hardest: Tressoir explicitly defers the “advanced form” knowledge-graph ontology as future work — and that future work is what Alhazen already is. The linked plain-text approach won because it’s easy for both humans and models to read, edit, and extend. If we realize the IB pattern graph-natively in Alhazen, we must not lose that legibility.
Glia & Engram (linked MIT talks)
Two talks from the same MIT Networks & Mobile Systems lab; Engram is the direct sequel to Glia. Glia is the agentic research loop; Engram diagnoses what Glia can’t do over long horizons and fixes it with a memory architecture. Read together, they’re a complete story about running an expert-level design loop and making it persist. Engram beating human SOTA on 8 of 9 systems problems is the strongest validation of the persistent-memory-over-context-window bet I saw.
GEPA / optimize_anything
(link) For me, this was the standout breakthrough of the meeting. The ARC-AGI jump — 32.5% → 89.5% by evolving an entire agent architecture as a text artifact in one declarative API call — was one of the most striking single results at the conference. GEPA beating GRPO by up to 20% with 35× fewer rollouts reframes the optimization question: language is a richer medium than scalar reward.
An Agentic analysis of the conference
I did apply my analysis methodology over the conference papers. I ingested each title, abstract, authors, and used agents to read every abstract and clustered them into ten thematic groups, tagging each one. The clusters aren’t categories assigned by the program committee; they’re the groupings that emerged from the abstracts themselves. The ten clusters: program optimization, agent memory and weight-free learning, multi-agent coordination, evaluation and benchmarks, safety and governance, serving infrastructure, coding agents, data retrieval and knowledge curation, human-agent interaction, and agent societies.
The complete paper-by-paper data, cluster assignments, and claim-evidence structure is available as a structured JSON export — the same representation that lives in the knowledge graph. If you’re interested, maybe download it and have Claude review it for you.
For a complete and global overview from the conference organizers, see the official write up here.
The full paper-by-paper dataset — 99 abstracts, 10 theme clusters, 13 tags, 3 meta-trend claims with paper-level evidence — is available as a structured JSON export: /data/cais2026-investigation.json. The knowledge graph backing this report uses the same alh-aboutness claim-evidence-source structure described in the Skillful-Alhazen project writeup.
CAIS 2026 took place in San Jose, May 26–29, 2026.