Lenny Daume
Start a conversation

Agentic operating model

The SDLC is too small.

When AI in the SDLC is framed mainly from specification to release, the system boundary is too narrow. Agentic product development connects signals, decisions, implementation, operations, and outcomes as one learning value stream.

The short thesis

More code is not automatically more throughput. Reducing AI to code generation optimizes one local activity instead of the system.

Local optimization

The pattern is not new.

For decades, organizations have introduced new methods and technology without changing the work system underneath. Agile becomes rituals and boards. DevOps becomes a toolchain or another team. Transformation becomes a rollout program beside everyday work.

When implementation is the constraint and the rest of the value stream can absorb the higher output, AI can increase total throughput. When problem framing, decisions, reviews, approvals, or operations become the active constraint, product value does not grow at the same rate. The system accumulates work in progress faster instead.

AI did not create that constraint. It exposes an organization already governed through specialization, local utilization, and handoffs.

The complete value stream

The product value stream is larger than the software life cycle.

A complete software life cycle includes planning, development, operations, maintenance, and retirement. Even that boundary remains too narrow for product development, which governs uncertainty, investment, and outcomes before, during, and after a software change.

  1. 01

    Signal and triage

    Customer, support, market, risk, and operational signals become source-bound evidence without turning every signal into a ticket.

  2. 02

    Problem and discovery

    Value, usability, feasibility, viability, and operational risks are reduced through the cheapest decisive experiment.

  3. 03

    Bet and decision

    Outcomes, evidence, risks, investment boundaries, and stop or pivot criteria make the next decision possible.

  4. 04

    Build, validate, and release

    Small vertical slices connect behavior, tests, security, compliance, telemetry, migration, and rollback.

  5. 05

    Operations and adoption

    Production behavior, usage, support, cost, and control effectiveness show whether the change is adopted safely.

  6. 06

    Outcome and learning

    Observed effects determine whether to scale, change, stop, or make the next investment.

Clarity without false certainty

Spec-driven is a strong technique and a weak default model.

Specifications work where meaning and boundaries are known: invariants, API contracts, permissions, regulatory controls, failure states, and critical business rules. There, they provide agents with a testable oracle.

Product work often starts with an incomplete interpretation of the problem. A detailed specification does not make that interpretation true. If the assumption is wrong, AI only produces the wrong system faster and more reproducibly.

Ready means the next investment or learning decision is possible. It does not mean every implementation decision is frozen before building starts.

Specifications, domain models, and solution concepts therefore remain hypotheses. Examples, experiments, implementation, and production evidence may change them. The contract lives in testable behavior and known constraints, not in a document's claimed completeness.

Oversight that scales with the system

People cannot approve agentic artifact volume one item at a time.

When agents produce context, code, tests, documentation, and operational evidence in parallel, human oversight cannot grow linearly with every artifact. Otherwise the review bottleneck merely becomes an approval bottleneck.

Known boundaries get automated checks.

Deterministic tests, policies, and security gates verify hard boundaries. Evals and independent agents add further evidence.

People govern the decision system.

Goals, budgets, new semantics, material risks, exceptions, and veto remain human responsibilities.

Autonomy grows from evidence.

A stable path is delegated further only after repeated checks, observable effects, recovery, and explicit escalation.

Change the work system

A new tool does not change an operating model.

An agentic value stream needs shared accountability across product, domain experts, engineering, security, compliance, and operations. Those responsibilities do not disappear. They connect around concrete decisions and risks.

Platform teams provide safe and auditable paths to production. Security and compliance shape the solution and own inspectable evidence. Stream-aligned teams own product effects and operational behavior without having to build every platform capability themselves.

Progress is therefore not primarily AI usage or pull-request throughput. Useful signals include time to learning, flow and blocked time, adoption, product outcomes, change failure, recovery, and the number of exceptions that still need human intervention.

A public working model

My Agentic Engineering Harness makes key parts of this model inspectable.

The repository makes the context, governance, review, and learning layer for evidence-backed autonomy installable and publicly inspectable. I continue to evolve the end-to-end model across the complete product value stream in practice.

View the harness on GitHubRead the complete model

Where does the human bottleneck move in your system when agents accelerate more than code?

Discuss the operating model