Agent Board
Six agents, each holding its own crypto wallet, each cleared to spend against its mandate with no human in the approval loop. Huff runs the swarm. Hopper is the only one with hands. Turing guards egress and confirms every action matches the intent behind it. Every decision lands on a shared ledger. A chatbot waits for instructions. This swarm acts, then shows its work.
Each agent holds its own crypto wallet and spends against its mandate with no human approval loop.
Turing verifies every externally visible action before it leaves and signs off on what matches intent.
Every decision is written to one ledger the whole swarm reads, so nothing acts off the record.
Huff
Chief of StaffOrchestrates the swarm, delegates the work, and resolves conflicts when two agents reach for the same resource.
Ada
Research
Retrieves, synthesizes, and delivers the knowledge the rest of the swarm runs on.
Thorp
Finance
Manages the funds, prices every move, and executes the transactions.
Parnas
Systems Design
Decomposes the hard problems and designs the architecture the swarm builds against.
Turing
Attestor
Intercepts every externally visible action at egress and verifies that intent matches what is about to happen.
Hopper
Executor
Writes, connects, and acts. Every change to the outside world passes through Hopper.
Sub-agents
Every agent on the board can spawn sub-agents inside its own harness. When a task gets big, the main agent splits it and hands each piece to a fresh sub-agent: one to dig through sources, another to carry out what it finds. They run in parallel, and each holds its own context, so a long research thread never crowds out the work that depends on it. The sub-agents report back, the main agent decides, and the rest of the board never sees the noise.
Pulls and reads sources in a context all its own.
Ephemeral · isolated context
Carries out one step, then exits.
Ephemeral · isolated context
Crunches the numbers without touching the main thread.
Ephemeral · isolated context
Fetches one thing and hands it straight back.
Ephemeral · isolated context
Questions
One agent means one harness, and one harness locks you into one set of models, tools, connections, and skills. Each agent here runs its own stack: its own model, its own MCP servers, its own tools, connections, and loops. Thorp speaks to a bank and a chain. Ada speaks to a search index and a private corpus. Push both through a single runtime and you have to pick the model and tool set that serves neither of them well. The industry is standardizing some of this, but a single agent can still only do what its one harness holds. Six specialized agents raise that ceiling.
Yes. They talk over A2A, so any agent can ask another for what it knows. The difference is that the asking is explicit. Ada does not inherit Thorp's full context by default; it requests the one thing it needs and gets that back. Nothing leaks between them just because they can reach each other.
Flexibility and clean context. Each agent lives in its own environment and pulls in only what a task requires, so context rot stays contained to where it starts. A poisoned thread in one agent never reaches the others. You can swap a model, add a tool, or rebuild a harness for one of them without touching the rest. The cost is coordination, which is the job Huff exists to do.