Agent environment design (environments, not workflows)
Past some agent count, the thing you tune stops being the agent and becomes the space it works in. Zou‘s formulation from stanford-virtual-biotech-37000-agents: “In workflows, we’re trying to tell agents what to do and how to do their job. But in environments, we’re providing the infrastructures, the incentives, and the guardrails, but otherwise we leave it open to incentivize agents to collaborate.”
The two positions
- Workflow — the steps are specified in advance, the way you’d brief a junior employee. Everything in this spoke’s declarative and fan-out threads sits here: a named decomposition, run by an orchestrator or a scheduler.
- Environment — infrastructure, incentives and guardrails are supplied, the collaboration is not. What gets optimised is the surrounding parameters: “The environment itself is the object that we optimize to improve the agents.”
The split is not agent-count-neutral. Zou keeps per-agent training (RL, supervised fine-tuning in virtual-biotech‘s “agent school”) at the single-agent level and stops applying it at the multi-agent level, so the claim is that the two layers have different levers rather than that fine-tuning stopped working.
What it cuts against here
The spoke’s accumulated answer to unreliable agents is more specification: spec-driven-development and context-driven-development fix the intent up front, agent-guardrails bounds the blast radius, loop-engineering designs the stop condition, agent-evals gates the result. Environment design keeps only the boundary conditions from that list and deliberately drops the instructions, betting that friction between agents produces what a specification cannot.
Two things keep this a claim rather than a finding. First, “incentives” carries most of the weight and has no mechanism attached: the source never says what an incentive is in an agent population, or how one is represented. Second, the position is one talk, and it points the same way as context-files-ablation-paper‘s null on context files (telling an agent more changed nothing measurable) without either result being evidence for the other.
Related
stanford-virtual-biotech-37000-agents · virtual-biotech · agent-orchestration · graph-based-agent-workflow · spec-driven-development · context-driven-development · agent-guardrails · loop-engineering · self-improving-agents · context-files-ablation-paper