Discovery, jurisdiction, and three signals from the frontier

Hey —

This week ended up orbiting two bottlenecks in the age of agents.

The first is discovery.

Once building gets cheap, the hard part becomes knowing what to build next. I wrote about that through hfpaper, a small CLI/MCP tool that gives agents a way to search, read, and cite research papers directly from the terminal. The tool itself is useful, but the deeper point is bigger: we’re moving from building to discovery. The build step keeps collapsing. The frontier moves to finding the next thing worth doing.

The second is jurisdiction.

Once execution gets cheap, misfiled context becomes expensive. In The Wrong Room for the Right Thought, I wrote about solving an issue from the wrong lane in my OpenClaw workspace — and what that revealed about agent systems. The issue got solved locally, but truth arrived globally only later. Other threads could keep investigating a problem that no longer existed. That’s a new tax: not execution cost, but context reconciliation cost.

I think these two are related.

One is about finding signal. The other is about making sure signal lands in the right room.

Three signals from the frontier

I also spent some time combing through the past week’s consumption. The pattern I see is not “general agents are winning.” It’s that the stack is quietly breaking into specialist primitives.

1. Search is becoming its own agent layer

Chroma Context-1 is a 20B search agent positioned as a retrieval subagent, not an all-purpose answer model. Open weights, cheaper, faster — but the real tell is conceptual. Search is splitting off into its own specialist layer.

That lines up with hfpaper. Discovery is no longer a feature buried inside a bigger model. It’s becoming a first-class surface.

2. Real-time voice is crossing from demo to deployable infra

Google shipped Gemini 3.1 Flash Live for real-time voice + vision agents. Ignore the launch copy for a second and look at the shape of the release: lower latency, better tool reliability, session management, ephemeral tokens, multilingual support.

That’s the difference between “voice AI demo” and “you can actually build a product on this.” The voice layer is hardening into infrastructure.

3. Memory is moving from optional tool to automatic context

Nicolò Boschi wrote about local long-term memory for OpenClaw using Hindsight. The important idea is not the plugin itself. It’s the architectural shift: memory works better when it’s injected automatically into context rather than exposed as a tool the model has to remember to call.

That matters because models are still bad at remembering to remember.

The alpha I see

The frontier is not one giant “agent” swallowing everything.

It looks more like a stack of increasingly sharp specialist layers:

• one thing for discovery
• one thing for search
• one thing for voice
• one thing for memory
• one thing for coordination

That decomposition is where I’d pay attention.

The people who win this phase may not be the ones building the biggest generalist. They may be the ones building the best primitives — and knowing how to compose them into something that feels coherent from the user side.

— Zak