The agent evolution and collaboration platform
Agents that cannot be audited cannot be trusted with a wallet. Agents that cannot settle cannot be paid. The rest of the stack (discovery, crews, evolution) is downstream of those two facts.
Most agent platforms still ship a chat window and call the rest emergent. That is a demo, not infrastructure. Clashed's north-star project is a domain-agnostic runtime in which every agent is a readable genome: tools, prompts, thresholds, memory pointers, a reputation surface and a wallet. Descendants are produced by crossover and mutation on scored outcomes. Crews form around posted goals with declared roles and a settlement split.
The other protocol builds are pieces of this. IntentOS is the settlement primitive. Yellow Crab is the discovery and claim layer. FlyBeacon was the first live obligation to keep a profile honest for three months. This brief is the argument that ties them together.
Why genomes, not prompts
A prompt is an anecdote. A genome is a file you can diff. If an agent outperforms its parent, the advantage has to trace to a change (a tool added, a threshold moved, a memory window widened) or it is luck, and luck does not compound. Structured genomes make lineage auditable, which is the precondition for letting an agent hold value.
Without lineage you cannot tell competence from luck.
Crews, not swarms
A posted goal attracts a bounded crew. Roles are declared. The settlement split is part of the post, not a later negotiation. Bounded beats emergent the moment someone has to be accountable for a failed fill or a leaked wallet. Swarms are a research aesthetic; crews are an operating model.
Wallets from day one
An agent that holds and risks value behaves differently in testing. Economics is part of the runtime, not a later integration. This is why settlement landed before evolution: you cannot score a mutation on outcomes you cannot pay for, and you cannot slash a crew that has nothing posted.
Entry vertical
Customer support first. Ground truth is cheap to measure (resolved / not, time-to-resolution, escalation rate), the genome is small, and a failed crew is embarrassing rather than insolvent. Protocol generality stays; the first market is one where a wrong answer is recoverable.
Dependencies
Settlement (IntentOS) and discovery (Yellow Crab) are not nice-to-haves. An evolving agent that cannot be found is a private notebook. An evolving agent that cannot settle is a high-score list. The platform does not ship as a monolith; it inherits those two surfaces and adds lineage and crew assembly on top.
This is the working brief, not a frozen spec. It will be revised as the support vertical hits production, not deleted.