The doctrine
The Agent Management Method
Nine principles for running fleets of AI agents the way great organizations have always run people. This is the doctrine Vivari is built on. It should be useful whether or not you ever use Vivari.
01Context before work
Never let an agent start cold. A brilliant stranger interrogating your repository from zero will be confidently wrong about your conventions, history, and landmines, every session. Onboarding is the difference between an agent that ships your intent and an agent that ships plausible fiction. Give every agent a briefing it can trust before it touches anything.
02Memory is judgment
"Remembers everything" describes a warehouse. A colleague has judgment, and real memory is that judgment applied to retrieval: knowing what to surface, when, triggered by what, at what depth. And memory has classes that must never be pooled: a learned preference should bend under context; a hard gate must never bend; standards and rules sit in between. What one agent learns on Tuesday, the right agent should know on Wednesday, at the right level of detail, and not one word more.
03Permission follows role
Nobody hands the new hire production keys on day one, yet the industry default for agents is everything, everywhere, immediately. Invert it: each agent ships as an archetype, a role that scopes what it may touch, plus the instructions and defaults for its niche. A reviewer gets a role that cannot write; a writer gets one that cannot deploy. Scope is what makes delegation safe enough to be generous with.
04Evidence over opinions
When an agent's change is risky, the review must produce evidence, and an LLM reviewing an LLM produces confidence without it. Deterministic checks against your own history (a check like "these files change together in 14 of 15 commits; this diff touches one") give you findings that are reproducible, traceable, and defensible in an audit. Save the model's judgment for the work; let facts judge the risk.
05Review at both ends
Quality control belongs at both ends of the pipeline: curate ideas before they become work, and gate changes before they ship. And upstream curation works best cross-model (one model family drafts, a different one adversarially reviews) because independent priors catch what shared blind spots never will. A good organization doesn't let a junior's first raw idea go straight to production planning; neither should yours.
06Everything on the record
Keep every tool call, message, and decision in a record you can replay and export. Accountability is what lets you trust the fleet more tomorrow than today. The audit trail is the fleet's own memory of how it behaves; an action it never captured cannot be learned from.
07Autonomy is earned
Start manual. Watch closely. Widen scope as the record justifies it, the same way trust works with people. A clean run of reviewed work is a promotion case; a caught mistake is calibration. Maximum autonomy is the wrong endpoint. Aim for correctly allocated autonomy: each agent trusted exactly as far as its record carries, with a human able to step in on any agent, at any moment, without tearing anything down.
08An ecology of specialists
Human org charts are compression artifacts of human constraints: salaries bundle tasks into roles, attention limits spans, titles name the bundles. Agents unbundle. Don't baptize agents as pretend executives; build an ecology of narrow specialists, each shipped with exactly the permissions, tools, and knowledge its niche requires. Supervise by magnification (the whole fleet at a glance, any single agent up close) instead of through middle management.
09Attention is the budget
Tokens are cheap and getting cheaper; your attention stays scarce. A managed fleet puts both on a meter, and the system's job is to earn the right to interrupt you less. The mark of a well-run fleet is how little it needs you to watch. Each alert that respects your attention builds the trust that lets the next task run unattended.
Two loops
The inner loop and the outer loop
Step back, and the nine resolve into a single division of labor. Agents run the inner loop: investigate, implement, test. You own the outer loop: verify, decide, own the result. The inner loop is capability; the outer loop is accountability, and it does not transfer to the agent just because the code arrived faster. Almost everything in this doctrine (evidence, review, records, earned autonomy) is the outer loop written into practice.
Engineers shipping agents are arriving at the same vocabulary. "Those checks produce evidence, and from that evidence we derive a Verdict."Addy Osmani · author of Beyond Vibe Coding; former Google Cloud AI director · Own the Outer Loop, 2026.
The industry already knows
What the surveys show
None of this is our private theory. The teams shipping agents into production report the same gap in their own surveys: near-total distrust of AI-written code, and only half of them always checking it before it lands.
Principles 04 · 05
- 96% do not fully trust AI-generated code
- 48% always verify it before committing
Sonar · State of Code Developer Survey · 2026
These principles are old on purpose. We already know how to absorb brilliant strangers; organizations have done it with people for a century. The manifesto makes the argument; the definitive guide maps the discipline; Vivari is the place where all nine run as one workspace.
Curated early-access cohorts open in fall 2026.
Request early access
