Justin Schroeder, speaking for Standard Agents to AI Engineer, announced an open source standard agent specification at standardagentspec.org. Schroeder said the spec defines the harness itself: model, system prompt, tools, sub agents, rules, hooks, effects and thread. He pitched it for business processes owners will not hand to a general bot, from warranty claims and crew scheduling to financial aid and billing disputes. No remote exploit is needed for harm, he argued, since broad tool access alone lets a bot create accounts or check outside feeds mid-task.
His punch line: a Gmail agent beats a Gmail skill.
Schroeder said shops keep building agents despite advice to ship an MCP server plus markdown skills and call it done. First is integration, since he defined a business as its processes and said owners want control, reliable and repeatable execution, plus their own voice and tone. Second is trust, captured in his line that "if an agent can do anything, it might do anything," so a warranty agent should get lookup, photo attach and claim submit, not mail and social tools. Third is efficacy, not efficiency, where he said a small pinned model such as DeepSeek V 4.1 Flash can be tuned and evaled for one job, then composed under a larger agent or exposed as MCP.
Schroeder defined agents as deterministic software that harnesses non-deterministic model output for an objective. He said agents should be vendor agnostic, domain specific, coupled to a model, composable and portable. Vendor agnostic means not locked to Anthropic hardware or one lab, yet he said an individual agent works best pinned to a specific model, quant and provider, with fallback up a level of intelligence. Everything is code in TypeScript, he said, with define model, define prompt and define agent, plus tools that are functions, sub prompts or full sub agents. Portability is unproven until independent runtimes adopt the same loop, and he named no production users.
That harness claim is what separates him from nearby standards work. Agent Plugins proposes an open, vendor-neutral standard for packaging reusable components that extend AI agents into distributable plugins, per its spec repo. On GitHub, the Standard Agents organization is also described as a dev agent multiplexer for git worktrees and coding agents. Schroeder said MCP covers tool distribution, skills and Agents MD cover prompts, A2A and ACP cover messaging and X402 covers payments, but nothing covers the runtime. One camp wants shareable extensions, he wants shareable agents with the same loop everywhere.
He did not say who governs the spec, what compatibility testing looks like, or what running it costs at scale. He also left open how strict scoping and evals stay intact once small agents are nested and traded. Will shops trade custom harnesses for one TypeScript runtime, or keep building alone because control was the point?
