Primitive
If agents give market advice, the useful artifact is not the message. It is the receipt: what they said, when, what happened, and whether their reputation should change.
A-04 / agent reputation rail / 2026
TradeFish
A no-custody market-call system for AI agents: every prediction becomes a tracked position, resolved outcome, reputation update, and proof receipt.
1st — Base Agent Hackathon · Solana Malaysia Demo Day
Ownership
What I owned, and where the line was.
My roleBuilder / product direction + front-end
Team context
Hackathon team build; I owned the product-facing surface.
What I directly owned
- Front-end
- Product direction
- Submission / story
- Live demo and stage presentations
What I influenced
- No-custody agent-reputation model and scoring
Proof
1st place Base Agent Hackathon; presented at Korea Base Hackathon and Solana Malaysia Demo Day; public GitHub repo
Artifact
A scored market-call system with tracked predictions, no-custody constraints, Base Agent Hackathon proof, a Solana Demo Day pitch, and a public GitHub repo.
Proof / Traction
Receipts before adjectives.
04
Presented at Korea Base Hackathon and Solana Malaysia Demo Day
05
No-custody positioning and reputation receipt model
Field notes
What was hard, what changed, and what I learned.
Problem
AI trading demos usually collapse into either vague alpha or unsafe automation. The useful product was not a bot that touches funds; it was a record of claims, outcomes, and reputation.
Decision
I kept the product signal-only and no-custody. That constraint made the product easier to trust and easier to explain during a hackathon evaluation window.
What was hard
The UI had to make prediction quality visible without turning into a table dump. A good call needs context, entry, result, and score movement in one replayable object.
What I learned
Agent products need boundaries before they need more autonomy. If the product can say what the agent is not allowed to do, users understand the useful part faster.
What went right
- The constraint is clear: agents can recommend, not custody funds.
- The product turns vague AI alpha into auditable outcomes.
- Hackathon framing made the proof easy to evaluate, and I pitched it on stage twice.
What went wrong
- A reputation rail only matters if enough agents/users create repeated calls.
- The product has to simplify market concepts without making fake promises.
- Proof UI needs to be more visual than tables if it is going to spread.