On Aug 1-2 I went to Agentic AI Summit 2026 in Berkeley. I have a 12 page report generated by AI from the raw notes and pictures I’ve taken, but that’s not what I want to share. This post is a few practical suggestions and a handful of takeaways / general observations.

First, some general conference attendance learnings:

• if you go somewhere for a conference, spend the time on the face-to-face as much as possible. The talks and materials will be online. The poster presenters standing next to their work at 2pm, who will answer any question you ask, will not be. On the second day I went around and talked to nearly all of poster presenters, and it was the best few hours of the two days.
• the people who are giving the talks are real humans, with profiles on LinkedIn and twitter, following them and connecting gives that mental bridge that connects the dots in your brain. Plus, as someone who has publicly spoken before, receiving LinkedIn connection request “say your talk” is great.
• however silly this might look like, take pictures of the slides, I found that the act itself of taking a picture and then looking at them later gives a way of recalling the information. And as a major perk, you can just throw all of the pictures into your AI agent and it will create the summary of what what happening.

Second, some observations, takeaways, and very quick thoughts:

• AI can be compared to fire as a general purpose technology. It took centuries to make safe and extremely efficient use of fire but now it is ubiquitous and same will happen to AI. While this comparison is heavily used and is nice, in my opinion, it might be wrong to compare because the fire had nothing to do with replacing or enhancing human cognition.
• Handling long horizon contexts and extensive work seems to be an overarching problem the industry is struggling with. Everyone is building some kind of special workflows around this. Workflows are nice, but they seem to be purely workarounds to fix LLM shortcomings (limited context, validation, etc). This might be the next major problem to solve. I could imagine LLM that is capable of constructing workflow and then following it, instead of you building that workflow for LLM.
• Many big companies, use all the same models from just a few major AI players on the market, what makes the difference is what companies build around the models.
• We have moved from tokenmaxxing to efficiency. A lot of talks were dedicated to reducing token waste. Entire startups exist purely to increase efficiency of model usage.
• Way too many talks and posters were dedicated to building benchmarks and comparing models, some of that was a bit repetitive, boring, and likely obsolete very soon (just being honest here).
• Some say that the harness is the secret ingredient and is the key differentiator in performance. On the other end, others say that whatever functionality you build into harness will be eventually learned by models. I personally don’t know how this will evolve, but I think 2 parts will exist for long time: one – something operating on environment like harness, and, two – the model itself and the responsibilities between the two will be fluid.
• Frontier labs often move fast, in 6 month something they built might be gone, but usually they manage to reap the benefits of that work in short time, thus allowing them to further invest and iterate and thus advancing entire industry.

All in all, I think it is useful to attend conferences in your field of work in order to learn where the industry is heading, what is on people’s mind, what companies are investing into, and most of all to have face-to-face time and networking. Materials presented are oftentimes available through other means, so in a way the material is just the excuse that gets everyone in the same room.

Thanks for reading.

If were were more interested about the content of the summit check this Dawn Song’s tweet.