The hype cycle wants you to believe more agents equals more capability. Stratum Praxis argues the opposite: adding agents increases decisions, permissions, handoffs, context boundaries, and potential failure paths as well as capacity. Before you scale your swarm, you need explicit contracts defining purpose, permitted tools, allowed decisions, and stop conditions. Autonomy without these boundaries is just chaos with a dashboard.

The Agent Contract as Firewall

You cannot grant autonomy until you define the cage. An agent contract can define purpose, permitted tools, allowed decisions, and stop conditions before more autonomy is granted. Human gates are most useful around consequential actions such as publishing, deleting, paying, contracting, or changing production state. If your agents can’t be supervised reliably within these bounds, they shouldn’t be running at all.

Evidence Over Internal Completion

Stop treating internal completion as revenue. Evidence makes agent behavior inspectable by recording what action happened and what result came back without treating internal completion as revenue. Context should be routed to the role that needs it instead of being broadcast to every agent by default. The useful scaling question is whether bounded roles can be supervised reliably, not how large an agent count can be displayed.

A Six-Point Audit for Current Workflows

Before buying or building new agents, use this decision aid on one existing process. First, choose one real recurring workflow and record its trigger, current steps, owner, and the outcome that would count as success. Second, record actual inputs and permission boundaries, identifying which data or approval must not be transferred to an automated service. Third, record the current cost, time, error, or rework burden using measurements you already have; write UNKNOWN where evidence is unavailable.

Test Small, Reversible Changes

Decide which one change is small and reversible. Do not reprice or rebuild a working commercial system to solve a traffic problem. Test that change with a real user or an observable work event; distinguish their action from a click, checkout, purchase, and payment. If evidence is not yet available, keep the current route healthy and note the precise event that would justify another change.

Make One Bounded Decision

Continue if the existing workflow is adequately controlled. Improve one measurable step if a real failure is observed. Escalate when permissions, sensitive data, or material external effects require a specialist or owner. Avoid expanding a project simply because another tool or model is available.

Key Takeaways

  • Adding agents increases failure paths as much as capacity.
  • Human gates are required for consequential actions like paying or deleting.
  • Context routing beats broadcasting; send data only to roles that need it.
  • Scale is about reliable supervision of bounded roles, not raw agent count.
  • Never expand a project just because a new tool or model is available.

The Bottom Line

Stop chasing agent count and start auditing contracts; if you cannot supervise bounded roles with explicit evidence and human gates, you are just automating chaos.