From MBA to MAA: Structuring an Agent‑Native Company

We’ve had a century to perfect the MBA playbook for human‑only organizations. Now agents are joining the org chart. This post introduces a complementary lens: the MAA — Master of Agent Administration — a pragmatic way to organize teams, systems, and governance when autonomous and AI agents become first‑class colleagues.

If MBA asks “How do humans run the business?”, MAA asks “How do humans and agents run the business together, safely, measurably, and profitably?”

What is MAA?

MAA stands for Master of Agent Administration. It’s a set of principles and operating patterns for agent‑native companies:

• Design for agents as durable actors with owners, identity, and scope
• Treat prompts, tools, and policies as product, not glue
• Make safety, observability, and provenance default‑on
• Optimize for outcomes: cycle time, quality, cost, and trust

Why now

Agents can already execute real, bounded workflows (support triage, data QA, marketing ops, basic engineering chores). The bottleneck is no longer raw capability; it’s organizational design. Without clear ownership, guardrails, and measurement, agents either create shadow automation or stall in pilots.

Agent‑Native Org: Roles and Responsibilities

Start small, but name owners. A minimal RACI looks like this:

• Agent Product Manager (APM): Defines problem, KPIs, data contracts, and rollout
• Agent Reliability Engineer (ARE): Builds, tests, deploys; owns runtime, tracing, SLOs
• Policy & Safety Lead (PSL): Designs safeguards, reviews incidents, handles approvals
• Data Steward (DS): Ensures data quality, access, retention, and lineage
• Domain Sponsor: Signs off on scope, budgets, and outcomes in the business unit

Architecture: The Four Layers

Think in layers to avoid spaghetti:

1. Intent and Policy: goals, constraints, compliance, approvals
2. Planning and Reasoning: decomposition, tool selection, memory, critique
3. Action and Tools: connectors, functions, RAG, transactions
4. Observability and Control: tracing, metrics, cost, drift detection, human‑in‑the‑loop

You wouldn’t let a new human hire ship prod without access rules, reviews, and dashboards. Don’t do it with agents.

Lifecycle: From Idea to Incident‑Free Ops

1. Define the job: task boundaries, success metrics, failure modes
2. Simulate: replay real data, adversarial prompts, and policy edge cases
3. Pilot: gated scope, staged environments, kill‑switches, feedback loops
4. Operate: SLOs, on‑call, budgets, continuous evaluation, drift alerts
5. Improve: post‑incident reviews, dataset curation, tool hardening

Guardrails and Governance

• Identity: stable IDs, signed actions, audit trails for humans and agents
• Authorization: least‑privilege tokens, time‑boxed scopes, just‑in‑time elevation
• Policy: allow/deny lists, redaction, content and action policies baked into prompts and tools
• Approvals: human checkpoints for irreversible and costly operations
• Provenance: attach intent, inputs, model/tool versions to every action

Metrics That Matter (beyond latency)

• Task success rate and human‑override rate
• Time‑to‑completion vs. human‑only baseline
• Cost per successful task and per avoided error
• Incident rate and mean time to restore (MTTR)
• Data quality regressions and drift

A Starter Operating Model by Stage

Early teams don’t need huge headcount — just clarity.

Stage 1: One Agent, One Team

• 1 APM, 1 ARE (shared infra), 0.2 PSL, 0.2 DS, 1 Sponsor
• Scope: one well‑bounded workflow with clear success metrics

Stage 2: Agent Portfolio

• Central platform team (2–4 AREs) with shared tooling, tracing, and policy
• Multiple APMs embedded with business units; PSL and DS as part‑time guilds

Stage 3: Agent‑Native Org

• Formal Agent Platform (routing, evals, feature flags, cost management)
• Policy as code, red‑team function, and agent on‑call rotation

Capability Map

The table below is a simple MAA capability checklist to track maturity across layers.

Practical Tips

• Start with one painful, repetitive workflow with clear guardrails
• Use test fixtures and replays; never evaluate with anecdotes alone
• Make every agent action explainable, revertible, and attributable
• Separate data read from write; gate writes behind approvals and tests
• Treat prompts, tools, and datasets as versioned artifacts

Closing: The MAA Mindset

The MBA taught us to scale human organizations. The MAA helps us scale human‑agent organizations. Give agents identity, owners, scope, and SLOs; give humans control, transparency, and outcomes. Start small, measure honestly, and let results earn trust.

Get involved

• Repo (seed; will move to the foundation): yamz8/maa-foundation
• Feedback & ideas: Open an issue