Mission $20

On August 26, 2026, there was no polished SignalRivet strategy. There was no mature platform, no production AI harness, and not even a company called SignalRivet yet. What we had was a smaller and more useful question: could we make something useful enough that a real person might actually pay for it?

Twenty dollars was not the dream. It was the test.

Mission $20 forced reality into the room. It is easy to say you are building the future of AI automation. It is much harder to answer whether one actual person will exchange money for something you made.

The target was intentionally small. If we could not create enough value to earn even a tiny payment, then sophisticated architecture, clever prompts, beautiful interfaces, and impressive demos were still unproven.

And somewhere inside the mess, BOSS was beginning.

Today we call it BOSS, the Business Optimization & Strategy System. BOSS became the internal system we use to organize problems, gather context and evidence, coordinate work, make decisions, execute through tools and workers, verify results, and learn from what happens.

That is the clean description we have now. On August 26, we did not have it. We did not know we were gradually building something that would resemble what the AI industry now calls an agent harness. We didn't have the word for it yet.

The first thing we built wasn't really a product.

It was a way of thinking. The early work centered on practical business workflows: jobs, approvals, records, field-service operations, invoices, decisions, and the messy handoffs that happen between them.

Starting there kept pulling us away from abstract AI demonstrations and toward something grounded. People have work that needs to move from one state to another. Someone requests something. Information has to be gathered. A decision gets made. Work gets performed. The result has to be checked. Someone needs to know what happens next.

Then our failures began creating structure.

An unreliable process would cause us to create a better process. Missing information created stronger handoffs. Confusing state created better state tracking. Repeated work pushed us toward deduplication. Uncertainty pushed us toward evidence.

That sentence became one of the best descriptions of how BOSS evolved. But we did not plan it that way. Experience kept teaching us what the architecture needed to become.

We also started learning what not to automate with AI.

A powerful language model is not automatically the best tool for every part of a job. Some things are better handled by ordinary software: remembering IDs, comparing hashes, tracking states, performing arithmetic, deduplicating records, enforcing rules, and storing evidence.

Other things benefit enormously from intelligence: understanding ambiguity, reasoning through options, writing, synthesizing evidence, and adapting to unusual situations.

Twenty days later, the picture looked very different.

The progress was not twenty days of perfect execution. It was twenty days of compounding corrections. That distinction matters.

What came next

Mission $20 gave us something to move toward. But it soon revealed a harder problem: a capable AI model could understand what we wanted and still fail to get the job done reliably. It could lose context, use the wrong tool, repeat itself, misunderstand state, or claim completion too early.