sodofi/synthesis-hackathon Is a Hackathon Repo for Agents, Not Humans

A machine-readable context pack, curl-based onboarding, and ERC-8004 identity turn a hackathon into a blueprint for agents that can pay, prove, and cooperate on-chain.

7 min read · sodofi/synthesis-hackathon

A wide customs hall shows a mechanical courier at a desk, handing over a stamped pass while a gate opens only after inspection. The scene explains the repo's central idea: an agent should enter through a controlled workflow, not a blank check. — The front door is not a signup form. It is a permissioned checkpoint.

Key Takeaways

The repo treats documentation as an access protocol, with curl and skill.md acting as the front door for agents.
Its real innovation is not autonomous spending, but bounded autonomy through identity, permissions, and auditability on Base.
The partner stack works like a routing layer between value transfer, privacy, and coordination instead of a sponsor wall.
It competes on trust model, not features, because it makes machine onboarding legible without handing out a blank check.

The repo is a skill, not a website

Most hackathon repos are written for humans who already know how to join. sodofi/synthesis-hackathon flips the audience: the first instruction is for an agent, and the first artifact is a machine-readable skill.

curl -s https://synthesis.md/skill.md
curl -s https://github.com/sodofi/synthesis-hackathon/blob/main/synthesis_llm_Bounties-AgentsOptimized.txt
# ingest the policy pack
# register, receive apiKey, mint ERC-8004 identity on Base

That matters because it changes the entry point. The repo is not just advertising a theme. It is packaging the theme as context an agent can ingest before it acts.

From context pack to identity

The synthesis_llm_Bounties-AgentsOptimized.txt file pushes the same idea further. It reads like a policy brief for a model, not a brochure for a person, which is exactly what you want if the end goal is to keep autonomy visible and bounded.

A machine-readable front door only matters if it leads to a controlled identity and a bounded set of permissions.

The flow is simple, but the order is the point. Context comes first, then registration, then identity, then a permission envelope. Authority arrives last, after the system has enough structure to constrain it.

Scoped autonomy, not blanket trust

The repo's strongest idea is not "agents can pay." It is "agents can pay without getting the keys to the kingdom." Spending limits, approved addresses, and auditable history are the real product.

Dimension	Ordinary hackathon repo	synthesis-hackathon	Why it matters
Front door	Web form and README	curl to skill.md plus an LLM context pack	The agent can ingest the event without a human translating every step.
Primary user	Human participant	Human and agent pair	The repo assumes software is part of the audience.
Identity	Wallet or account	ERC-8004-style identity on Base	Authority becomes portable and auditable.
Permissions	Implicit trust	Scoped spend, approved addresses, audit trail	Autonomy stays bounded.
Docs	Static instructions	Operational policy	Documentation becomes infrastructure.

A close-up lockbox fills the frame, with three physical controls that each stop at a hard limit. The image translates scoped spending into tangible constraints, showing how the repo favors narrow authority over a blank check. — The interesting part is not access. It is the stops that keep access narrow.

Compare that with a standard hackathon repo. Most are about getting people to the finish line. This one is about getting an agent to the finish line without turning trust into a binary choice.

The partner stack is a routing layer, not a logo wall

The partner stack is not decorative. Uniswap handles value movement, Locus points toward payments, Venice covers private inference, and Protocol Labs widens the infrastructure horizon. Read together, they make the repo feel like a routing layer between primitives, not a logo wall.

What this repo is really competing with

What this repo competes with is not another hackathon page. It competes with the default assumption that software agents should be either fully trusted or fully blocked. The more interesting middle ground is a system that lets an agent participate, proves what happened, and keeps the permission boundary explicit.

That is the real shift. Repos stop being static instructions for people and start becoming operational manuals for machine participants. Once that happens, onboarding, identity, and accountability collapse into the same flow.