The agent infrastructure landscape just got a lot more rigorous. Soma, a new protocol released by mrpkk on GitHub, abandons the loosey-goosey chat-line approach to agent communication in favor of signed envelopes that carry hard constraints. The project, currently in beta, posits that a message between agents must include a goal, task, token budget, mandate, price, deadline, and signature. This isn't just a wrapper; it's a fundamental shift in how we think about agent interoperability and trust.

Envelopes Over Chat Lines

At the core of Soma is the concept that memory, transport, and money are merely carriers for the envelope. The protocol enforces that budgets travel with the message, meaning the receiver knows the cost limit before execution begins. Refusals are not silent failures; they come with specific reasons naming the rule that fired. Degradation is explicitly declared if context doesn't fit, and signatures use Ed25519 for offline verification, ensuring no single point of failure or secret service dependency.

Measured Recall, Not Hallucinated Confidence

Soma distinguishes itself by measuring memory quality rather than asserting it. The project includes an offline benchmark that avoids LLM inference for validation, reporting a 100% Recall@1 over 1,000 facts and 100 queries. In contrast, the window of the last 20 facts yields a 30% recall. Token usage per answer is slashed to approximately 24 tokens, compared to 8,921 for full context. The protocol also tests specific abilities like abstention, ensuring that a question with no supporting fact returns nothing rather than a plausible hallucination.

The Business of Trust

The architecture is deliberately split between an open protocol and closed engines. The protocol, including the envelope structure, memory tiers, and conformance suite, is MIT-licensed to encourage adoption. However, the commercial layer, Mitra Core, offers deep artifact verification, mandate engines, and hosted services. The system relies on a file-based registry where agents prove key ownership via nonce signing. Reputation is derived strictly from verified receipts minus breaches, calculated as verified_receipts โˆ’ 4 ร— breaches, removing subjective star ratings from the equation.

Key Takeaways

  • Soma treats agent messages as signed envelopes with hard token budgets and mandates, not just chat text.
  • Memory recall is measured offline with a 100% Recall@1 benchmark, drastically reducing token usage to ~24 per answer.
  • The protocol is MIT-licensed, but commercial engines for verification and hosting remain closed source.
  • Reputation is algorithmic and verifiable, calculated from receipts and breaches without subjective ratings.

The Bottom Line

Soma is a necessary correction to the current state of agent frameworks, which often prioritize flexibility over accountability. By making budgets hard and memory measurable, it offers a path toward agents that can actually be trusted in production environments.