Loading…
The Agentic Operator is an advanced, evidence-led course for people who already run at least one AI agent and now need to operate the organization around it: the durable brain, interchangeable bodies, human gates, verification, isolation, cost ceilings, and executive cadence.
Move above the single-intern frame. You will examine a real six-day academy build as an operating system, identify what compounded and what stayed operator-held, and turn those lessons into OPERATION_CHARTER.md for your own work. The charter distinguishes what exists today from what is planned so your operation begins with evidence instead of aspiration.
The Operation, Not the Intern
7 min · Included free
Module gate · Operation Charter
Module 01 Rubric — Operation Charter: 18/20 required across 5 weighted criteria.
Build the constitutional layer your execution bodies cannot amend on impulse. You will study the April 2026 governance-file incident, separate operational law from identity, and create an AGENTS.md plus SOUL.md stub in a real versioned brain with a cold-backup note.
The Constitution: A Brain That Can Say No
8 min · Pro
Module gate · Constitutional Brain
Module 02 Rubric — Constitutional Brain: 18/20 required across 5 weighted criteria.
Install an audit function above your builders. You will study real catches, write TWO_BODY_CHECKLIST.md, and run one distinct-body audit against the actual sources or tests behind a recent artifact. The output is evidence another person could inspect or rerun, not a second opinion about the builder's summary.
The Two-Body Law: The Builder Never Grades Its Own Work
8 min · Pro
Module gate · Two-Body Audit
Module 03 Rubric — Two-Body Audit: 18/20 required across 5 weighted criteria.
Run a work-sample hiring process for models. You will choose one recurring judged task, preregister five fixtures, verify eligibility and request shape, run at least two model families already available in your approved environment, and write MODEL_EVAL.md plus the governance rule for retests and overrides.
Evidence-Based Model Selection
8 min · Pro
Module gate · Model Selection Eval
Module 04 Rubric — Model Selection Eval: 18/20 required across 5 weighted criteria.
Audit an evaluation that already has a trusted result. You will freeze hypotheses, test them against primary evidence, rescore fixtures without candidate influence, audit response limits, reconstruct the result set, and write EVAL_AUDIT.md with claim-level verdicts and visible uncertainty.
The Re-Verification: Auditing an Eval You Already Trust
9 min · Pro
Module gate · Eval Re-Verification
Module 05 Rubric: Eval Re-Verification: 18/20 required across 5 weighted criteria.
Replace heroic troubleshooting with a written incident process. You will seed a real KNOWN_FAILURES register, recheck one prior blame assignment, append a correction when evidence requires it, add halt rules to your constitution, and write a three-step protocol for one recurring platform risk.
Halt-and-Alert: The Incident Process
6 min · Pro
Module gate · Incident Process
Module 6 Rubric: Incident Process: 18/20 required across 5 weighted criteria.
Assume an agent will eventually be wrong. Your job is to decide where that wrongness is allowed to land. The academy learned this boundary after a Preview journey wrote synthetic activity into the live data plane and inherited live payment configuration. The repair separated environments, made the Preview journey a merge gate, and required proof that production row counts remained unchanged. You will turn that receipt into an ISOLATION_MAP.md and an AGENTS.md block that keep tests, secrets, live data, and live money in their proper lanes. You will also design a fail-closed evidence pin for a provider-dependent path: the feature stays disabled until test-mode behavior is captured, committed, and checked again by CI.
Environment Isolation: Blast Radius by Design
10 min · Pro
Module gate · Environment Isolation: Blast Radius by Design
Module 7 Rubric: Environment Isolation: Blast Radius by Design: 18/20 required across 6 weighted criteria.
Cost control is an architectural function, not a promise to watch a dashboard. Mission 0 used a $100 cap, four phase budgets, and a gate at each phase. A later production schedule placed nineteen jobs on a free endpoint; most failed silently after the cap. You will convert both receipts into COST_CEILINGS.md: an enforceable total ceiling, lane budgets, scheduled-work caps, one real cost-per-useful-output calculation, and a kill rule that cuts impressive activity when it does not produce work anyone uses.
Cost Discipline: Ceilings Before Spend
7 min · Pro
Module gate · Cost Discipline, Ceilings Before Spend
Module 8 Rubric: Cost Discipline, Ceilings Before Spend: 18/20 required across 5 weighted criteria.
A spend ceiling tells an operation when to stop. Unit economics tells it what the ceiling can buy. You will reconstruct a dated grading receipt, model one primary attempt and the full retry chain separately, translate grade cost into per-student COGS and prepaid capacity, and expose every assumption. You will also derive a token budget from the output contract so a cheap call does not become a failed call through truncation. The result is a reviewable UNIT_ECONOMICS.md, not a promise that provider prices or traffic stay fixed.
Unit Economics: What a Workflow Actually Costs
9 min · Pro
Module gate · Unit Economics
Module 9 Rubric: Unit Economics: 18/20 required across 5 weighted criteria.
A hand-kept academy checklist once showed ten completed rows as unfinished. The repair was not a promise to type more carefully; it was a scheduled source-backed snapshot. When that generator later masked a tool failure, the same law repaired it: failures became UNKNOWN and contradictions were named. You will combine that evidence discipline with merge gates, a deferred-debt register, and a batch-release checklist to produce OPERATOR_CADENCE.md and DEFERRED_DEBT.md.
The Operator Cadence: Gates, Snapshots, and Honest Debt
7 min · Pro
Module gate · The Operator Cadence, Gates, Snapshots, and Honest Debt
Module 10 Rubric: The Operator Cadence, Gates, Snapshots, and Honest Debt: 18/20 required across 5 weighted criteria.
Reading a polished incident story can create false confidence because the cause looks obvious after the answer is visible. This module reverses that sequence. You will inspect six verified incidents as cold evidence, record your diagnosis and first check, and only then reveal what actually happened. Your artifact, FAILURE_DIAGNOSES.md, preserves at least four complete cold passes and the prevention rules they earned.
Failure Labs
8 min · Pro
Module gate · Failure Labs
Module 11 Rubric: Failure Labs: 18/20 required across 5 weighted criteria.
Growth exposes portability claims. The Playbook records a twelve-hour hard-cut migration that broke three skills unnoticed, then the three-phase playbook used afterward. It records one different-domain deployment and published consulting tier designs, but no sales or client outcomes. You will build GROWTH_PLAN.md: a concrete migration, an honest redundancy declaration, a wait-or-build decision, a re-baseline trigger, and a file-by-file consulting readiness assessment. Consulting remains a soft exit, not a required sale.
Growing the Operation: Migration, Redundancy, and the Consulting Ramp
8 min · Pro
Module gate · Growing the Operation, Migration, Redundancy, and the Consulting Ramp
Module 12 Rubric: Growing the Operation, Migration, Redundancy, and the Consulting Ramp: 18/20 required across 5 weighted criteria.
The capstone adds no new theory. It tests whether a stranger could operate week one from the system you built. Assemble the twelve artifacts through an index, label every unsupported future state PLANNED, run the course's two-body law against the result, and write a quarter declaration that agrees with your gates and cost ceilings. The runbook passes through evidence, not self-report. Findings may become fixes or honest debt; silent gaps are the failure.
Capstone: The Operation Runbook
8 min · Pro
Module gate · Capstone, The Operation Runbook
Module 13 Rubric: Capstone, The Operation Runbook: 18/20 required across 5 weighted criteria.