Skip to content

ADLC — Agentic Development Life Cycle

The Software Development Life Cycle (SDLC) was designed for teams of humans. It assumes that every actor in the loop reads specs the same way, follows conventions, asks questions when unsure, and learns from the codebase over time. That model starts to crack the moment an AI agent joins the team — not as a helper invoked ad-hoc, but as a first-class participant executing tasks day-to-day.

ADLC — the Agentic Development Life Cycle — is what the SDLC becomes once you take agents seriously. Okto Pulse is built from the ground up to support this cycle.

SDLC assumptionWhat breaks with agentsADLC response
Specs are “best effort” contextAgents need explicit, machine-readable specs to act correctlySpec is a first-class artefact with required fields
Work items flow loosely through stagesAgents jump stages without gates if permittedStage transitions are permissioned, explicitly
Roles are human job titlesAgents need surgical permission grids, not titlesRole = resource matrix (read/write/gate by artefact)
“Ask the team” for contextAgents can’t meaningfully ask a Slack channelContext is linked in the artefact graph
Validation is PR reviewAgents can out-throughput any human reviewerDedicated validator role gates spec/task/sprint

Every unit of work traces to a spec. No “just try this and see” tickets. Upstream roles (humans, or agents in the Spec preset) own ideation → refinement → spec. Downstream executor agents can only act on specs that are approved.

Executor agents (human-driven, AI-driven, or mixed) pick tasks linked to specs. They never touch the spec itself — that would be a conflict of interest. MCP exposes the board so any MCP-aware agent (Claude Code, Cursor, Windsurf, custom orchestrators) gets the full API, role-permissioned.

The Validator role has exclusive rights to submit spec/task/sprint validations. An agent executing can mark a task “ready for validation” but cannot close it. A validator agent (or human) evaluates against the original spec and either passes or rejects.

Task → Sprint → Spec → Refinement → Ideation. The chain is enforced at the data model. Every shipped card is auditable back to the motivating idea. This isn’t governance theater — it’s what lets you ask “why did we build this?” six months later and get a real answer.

  • Pipeline model — 5 explicit stages, each with artefacts and gates. See The Pulse Pipeline.
  • Role system — 5 built-in presets + Custom, mapping the ADLC roles above. See Agent Roles.
  • MCP server — 150+ tools spanning the entire board, role-aware. See MCP Integration.
  • Local-first data model — SQLite on your machine. You own the trace graph.

Why “life cycle” and not just “workflow”

Section titled “Why “life cycle” and not just “workflow””

Workflows are task-by-task. Life cycles are about how intent becomes a shipped unit of work and how knowledge accumulates across many cycles. ADLC is the latter — it names the whole loop, not just the kanban lane.