Agentic Surfaces
Why do agents get all the fun?
A few weeks ago I wrote about Strategies on Canix402 and Brownie Bot as a proof that an autonomous treasury could talk to the network without surrendering custody. That note ended on a fairly simple resolve: keep building either way. Since then, the work has moved from proof into production, and a second question has started to sit beside the first. If we are serious about agent-native infrastructure, what does that infrastructure look like when a human is still the person who needs to understand it?
This update is about that shift. Brownie is already live, Ghillie is nearly ready to take its seat, and we have been trimming storage and retiring older bots while the newer ones settle in. Alongside the shipping, we have also been thinking about agentic surfaces — ways for people to take advantage of the same enriched x402 data that agents already consume — plus how Neon Forge packages those offerings, and what’s next for CompX and the COMPX token.
Dev Update: Brownie Live, Ghillie Incoming
Brownie Bot is now live and managing the CompX community treasury, using Canix402 and ZeroSignal as the rails it researches, plans, and acts against. That’s the outcome we aimed for when we open-sourced the steward in the first place: not a demo that looks convincing in a terminal, but a bot that runs against real balances, real positions, and real policy limits, with signing still gated behind deliberate choice and keys staying where they belong.
In parallel, Ghillie Bot continues to move in the right direction. It is successfully using Amarok to interact with Alpha Arcade’s prediction markets, which is a different kind of loop from treasury allocation but the same underlying idea — an agent that can discover, reason, and eventually manage positions without rebuilding the plumbing every day. The older Ghillie lineage, Nuckelavee, has been turned off. We expect the new Ghillie to be up and running this coming week, taking over prediction-market management from the previous bot on the same wallet so continuity stays intact while the implementation improves underneath.
There has been other infrastructure work as well. We have been cutting databases and unneeded data, moving toward light storage and memory on DigitalOcean Spaces rather than carrying around more state than the bots actually need. It is not the glamorous part of agentic DeFi, but it is the part that keeps these systems cheap enough and clear enough to run for real.
Agentic Surfaces
As we have shipped more of this stack, a fairly obvious question has started to bother me. Why do we only give agents the enriched x402 data? Is it simply because of the way they interact — because an agent can pay for a payload, parse a schema, and move on without caring that the payment was a tiny transaction? A human is not going to want to sign multiple micropayments over and over just to browse a website or poke at an API. The asymmetry is not in the value of the data; it is in the interface.
That is where the idea of agentic surfaces comes in. Imagine a Chrome extension that handles the agent side of the relationship: wallet access for the small x402 interactions, payment negotiation, and the quiet work of fetching enriched responses. Then, instead of leaving that intelligence stranded in a machine-readable format, the extension brings it into the browser by modifying and adding to the DOM in real time. A prediction market like Alpha Arcade, for example, could gain an overlay that highlights deeper context, surfaces details the page itself never shows, and gives a human more leverage without asking them to become an agent engineer.
The important custody boundary stays familiar. The extension agent would only need a very small balance to interact with x402, while everything else — the larger signing decisions, the trades that actually move meaningful capital — would still be decided by the human, separately, exactly as it works today. The agentic layer handles discovery and enrichment; the person remains in charge of consequence.
How We’re Packaging Agentic Offerings
Inside Neon Forge we have settled into a design system that keeps working for us. A private API holds the real intelligence. Caddy exposes x402 endpoints — both free and paid — and returns enriched data to agents that can discover and pay for it. Then we ship an open-source user-agent so humans have a starting point rather than a wall of protocol documentation. That pattern is what Brownie is for Canix, and it is the shape we keep returning to: infrastructure for machines, plus an inspectable bridge that makes the same capability usable by people who want to run it themselves.
The next agentic offerings may be less DeFi- and crypto-focused. Consumer shopping data is an obvious candidate for enrichment — not just comparison shopping or price consolidation, but recommendations with depth. GPU pricing is a concrete example: an agent that can go beyond listing numbers and actually reason about value, availability, and trade-offs. Company data is another. Imagine asking an agent to tell you about a company and having it pull a full set of incorporation details, public holdings, LinkedIn context, and related signals through x402 endpoints rather than scraping half a dozen tabs by hand.
One idea that sits high on my own list is developer dependencies. An agent that can take — or directly run — something like npm outdated, then actually go and read the release notes, look for known issues, and give a recommendation on whether you should upgrade, ticks a lot of boxes for me personally. It is useful, clearly agent-shaped, and exactly the kind of careful work people already wish something else would do on their behalf.
CompX DeFi and the COMPX Token
Our DeFi platform at CompX needs a roadmap, and so does the COMPX token itself. We are committing to providing a roadmap for the remainder of this year for both the platform and the token by the middle of August — the 15th, to be precise. I have a lot of ideas for working with the token, and we are making decisions about the CompX clAMM product as well. Expect more concrete detail in the coming weeks.
What’s Next
The near horizon is fairly clear. Ghillie should take the prediction-market seat this week while Brownie continues stewarding the community treasury in the open. Agentic surfaces — starting with the idea of a browser extension that pays for enrichment so humans do not have to — feel like the natural next place to pressure-test whether this stack only works for machines, or whether it can make people more capable too. And by mid-August, CompX and $COMPX token should have a written path for the rest of the year.
We are still early. The bots are live enough to matter, the packaging pattern is clear enough to reuse, and the human surface is still mostly a sketch — and that is fine. The interesting work usually starts at the point where the experiment stops being theoretical and starts asking what it is actually for.
I’m building either way.
