How agentic is your platform, and how do you know it's safe?
Autonomy is not a switch — it is a ladder with a governance record. Six levels, eight dimensions, one rule: every rung is earned per action class on evidence, and every rung is revocable.
The short answer
Six levels — manual, assisted, automated, agentic, governed autonomous, adaptive — measured across eight dimensions. Most organizations should target L4 selectively: governed autonomy where actions are reversible and high-volume, human approval where they are not.
The ladder
Manual
Humans operate everything. Tools exist, but every action starts with a person.
- Reasoning:
- None
- Action:
- Human, via CLI/tickets
- Identity:
- None for agents
- Memory:
- Tribal knowledge
- Policy:
- Documents and reviews
- Observability:
- Dashboards for humans
- Human authority:
- Total
- Autonomy:
- None
Assisted
Copilots help humans write and review — but humans initiate and perform every action.
- Reasoning:
- Suggestions only
- Action:
- Human-executed
- Identity:
- Human credentials
- Memory:
- None durable
- Policy:
- Documents and reviews
- Observability:
- Dashboards for humans
- Human authority:
- Total
- Autonomy:
- None
Automated
Deterministic workflows automate tasks. Scripts and pipelines do exactly what they were coded to do.
- Reasoning:
- None (predefined)
- Action:
- Pipelines execute fixed steps
- Identity:
- Service accounts, often shared
- Memory:
- Logs only
- Policy:
- Guardrails in code
- Observability:
- Run logs
- Human authority:
- Set at build time
- Autonomy:
- Bounded and fixed
Agentic
Agents reason and propose actions. A human reviews and executes the consequential ones.
- Reasoning:
- LLM planning, drafts proposals
- Action:
- Proposes; human executes prod
- Identity:
- Emerging per-agent identity
- Memory:
- Incident notes, embeddings
- Policy:
- Policy drafts + review checklists
- Observability:
- Proposals logged
- Human authority:
- Approves every action
- Autonomy:
- Read and recommend
Governed autonomous
Agents execute bounded actions through policy-controlled interfaces. Humans hold the irreversible decisions.
- Reasoning:
- Grounded in platform context
- Action:
- Executes via golden paths
- Identity:
- Scoped, attributable, time-bound
- Memory:
- Scoped platform memory
- Policy:
- Policy-as-code, fail closed
- Observability:
- Full decision + action traces
- Human authority:
- Production and irreversible acts
- Autonomy:
- Read → execute, per class
Adaptive platform
Agents observe, optimize, remediate, and evolve platform state continuously inside defined governance.
- Reasoning:
- Plans and reflects on outcomes
- Action:
- Executes and remediates
- Identity:
- Mature delegation chains
- Memory:
- Compounding organizational IQ
- Policy:
- Continuously tested policy
- Observability:
- Evals + traces drive tuning
- Human authority:
- Sets bounds; agents operate within
- Autonomy:
- Up to remediate, per class
How to read the eight dimensions
The dimensions mirror the reference architecture's planes deliberately: a level only counts if its planes exist. L4 "governed autonomous" requires real identity (scoped, attributable), real policy-as-code (fail closed), and real traces — not a model with credentials and goodwill. An organization is at the lowest level its weakest dimension supports, for the action class in question.
How you actually move up
- 1.Pick one action class. Not "the platform" — one thing, e.g. config refactoring or first-pass remediation.
- 2.Instrument before you automate. Identity, policy gate, traces — the L4 plumbing gets built at L2/L3, while humans are still executing.
- 3.Climb the autonomy rungs on evidence. Read → recommend → propose → execute → remediate, each rung earned on a clean trace history for that class.
- 4.Keep the irreversible rungs human. Some action classes never leave L3-style approval — that is the model working, not failing.
The demo is this loop in miniature: an agent at L4 for one action class (deploy a compliant workload) — proposing, validating against policy, converging, and stopping at a human gate for production.
Where organizations actually sit today
Honest read of 2026: most platform teams are at L1–L2 organization-wide, with pockets of L3. Vendors demo L5; almost nobody operates it, because L5 presumes the identity, policy, and evaluation infrastructure of L4 at scale. The gap between L3 and L4 is where the discipline lives — it is the difference between agents that suggest and agents that act with a governance record.
Quick answers
What is the agentic platform maturity model?
Six levels describing how much agency a platform can safely exercise: L0 manual, L1 assisted (Copilot-style help), L2 automated (deterministic pipelines), L3 agentic (agents propose, humans execute), L4 governed autonomous (agents execute bounded actions behind policy gates with human-held irreversible decisions), L5 adaptive (agents continuously optimize and remediate inside defined governance). Each level is evaluated across eight dimensions, from reasoning and action to identity, policy, and human authority.
Which level should an organization target?
Most organizations with mature platform practices should target L4 selectively, not wholesale: governed autonomy for high-volume, reversible action classes (config refactors, right-sizing, first-pass remediation), while consequential and irreversible actions stay at L3-style human approval indefinitely. L5 is a direction, not a destination — chasing it everywhere usually means skipping the governance that makes any autonomy safe.
How long does it take to move between levels?
Moving up is per action class, not per organization — and it is evidence-gated, not calendar-gated. An action class earns each rung of autonomy (read → recommend → propose → execute → remediate) on a clean trace history, and can be demoted instantly when behavior degrades. A well-instrumented platform can move one action class from L2 to L4 in weeks; moving the whole platform is a multi-quarter program.
Ready for the leap?
Partner with Adventure On The Wave to build governed, agentic platform capability — architecture, guardrails, and the human authority model to match.
A strategic initiative by Adventure On The Wave