Notes from the agentic frontier.
Write-ups on the unsolved problems behind the Labs — verification and provenance, proof-of-human, agent memory and world-state, computer-use, and game intelligence. Less polished paper, more lab journal: the thinking that turns into systems.
The verification frontier: agents act, but nothing checks what they did
The hardest unsolved problem in agentic systems isn't getting agents to act — it's verifying, version-controlling, and proving what they actually did. The next layer of value is a trust layer for agents, not more agents.
Proof-of-human for the agent era
As agents learn to play games, drive UIs, and pass behavioral checks, 'is this a human?' breaks everywhere. The durable answer isn't classifying behavior — it's modeling the generative physics of human motor control, which an AI can't fake by matching statistics.
Agent memory needs version control, not just retrieval
Long-horizon and multi-agent systems silently rot: a fact gets revised and a dozen decisions built on the old one are never flagged. Retrieval is solved; the unsolved problem is the dependency graph from facts to the decisions that consumed them.
Rules-grounded game masters: what AI Dungeon got wrong
A coherent, long-horizon AI game master is one of the great unsolved problems in agentic media. The fix isn't a bigger model — it's a neuro-symbolic split: compile the rulebook into a constraint engine and route only genuine ambiguity to the LLM.
Computer-use agents need ground truth, not screenshots
You can't safely run an agent on real software if you can't verify what it actually did. Screenshot-judging vision models pass agents that reach the right pixels via the wrong path. The fix is to read the OS, not the screen.
How to run an adversarial multi-agent idea search
A brainstorm gives you the obvious ideas everyone already lists. To find what's genuinely unsolved, the research itself has to be adversarial: generate broadly, then have independent skeptics try to kill every idea. What survives is the signal.