video - agents need more than chat - 2026-04
https://youtu.be/XNtkiQJ49Ps?si=qo24Lw2iiH0Rhptm
My Summary
... then your bottlenecks change.
While some domains are seemingly more verifiable than others, if you look at the work in the domain and do task decomposition on the actual work, you'll see that some of those tasks are easily verifiable and other are hard to verify. The easy to verify stuff can be automated, and the hard to verify stuff can rely on people. You still get a speedup from automation, particularly if the person can do bulk review because the agent does speculative execution then presents the person with a decision log and the user can steer, in bulk, later rather than getting pinged for every little decision, which would cause the agent to get stuck waiting on user input)
He also makes the point that natural language is a limited UX. For a very complex and deep tree of tasks, chat is not a good interface. You want a standard UI powered by AI that surfaces tabular data as a management interface (for that bulk human review) or documents with annotations for getting more detailed.
Notes
At 4:00 he talks about the verifiability of different domains, but actually within those domains the individual tasks that are part of the domain have different levels of dim - verifiability
At 4:00 he talks about the difference between there's two things there's dim - level of control and dim - level of trust. Then later he talks about how you can increase trust with guardrails
At 6:10, he talks about how you can't actually verify a contract in legal, but you can get a proxy for it
...
Ways of doing verification in legal domain
At 8:00 he describes generating a legal report as essentially a big tree of tasks or a directed acyclic graph - DAG
...
That tree of tasks, which needs to be carefully executed makes me think of Gas Town molecule
At 10:30 elicitation which is asking users what to do but then instead of doing that you should make a decision and put it in a decision log and then the user can undo that decision later. So the agent doesn't get stuck.
...
design pattern - speculative execution because we want to avoid agent waiting for user input
At 11 minutes, he shows a tabular interface for review and also a document.