Start with decision rights
Define what the system may decide, what it may propose, and what always requires human authority. Autonomy is a permission model, not a personality trait.
Stage 04 · Operate · 2026 field report
Production agent operating principles.
A practical synthesis for teams moving from compelling demonstrations to systems that can be trusted with real work.
Published 21 July 2026 · Agentic Systems Editorial Team
Executive summary
Capability creates value. Operating discipline makes it durable.
Production agents combine probabilistic decisions with consequential access to data and tools. That makes ordinary software controls more important, not less.
The strongest operating model keeps authority explicit, state inspectable, evidence continuous, and human intervention designed into the normal path. Teams should increase autonomy only where observed performance and recoverability justify it.
Six principles
Make success externally observable
Completion should be supported by tests, records, citations, or confirmed state changes—not by a fluent statement from the model.
Keep control deterministic
Authentication, authorization, budgets, schemas, idempotency, and irreversible-action gates belong in code and infrastructure.
Evaluate trajectories, not only answers
Inspect tool selection, intermediate state, recovery behavior, escalation, cost, and latency alongside final-output quality.
Design for graceful handoff
Uncertainty, missing authority, conflicting evidence, and exhausted limits should produce a clean transfer with useful context.
Treat change as a new claim
A model, prompt, tool, policy, data source, or workflow change can invalidate prior evidence. Re-evaluate in proportion to risk.
A simple maturity model
01 · Assisted
The system drafts or recommends. A person performs the consequential action.
02 · Supervised
The system performs reversible work and pauses at explicit approval boundaries.
03 · Bounded
The system operates within measured permissions, budgets, monitoring, and escalation rules.