Skip to content

CONCEPT Cited by 1 source

Agent Development Lifecycle

The Agent Development Lifecycle (ADLC) is a framing for software delivery in which agents operate across planning, implementation, testing, deployment, maintenance, and retirement, rather than serving only as code generators. Cloudflare proposes it as a successor to a human-managed Software Development Lifecycle when agent implementation capacity makes the downstream validation and operations path the bottleneck. (Source: sources/2026-08-04-cloudflare-agent-development-lifecycle)

What changes from an SDLC

An SDLC can automate individual steps while a human still advances work, interprets dashboards, applies feedback, and decides when to continue. The ADLC moves those handoffs into an explicit control loop. The goal is not unattended code generation. It is a lifecycle in which agents can safely receive a production error, customer report, or product idea, then create evidence and make bounded progress across the whole delivery path.

Required properties

The Cloudflare framing identifies seven conditions an agent-operated lifecycle needs:

  1. Programmatic: every operating action is exposed through debuggable APIs rather than dashboard-only procedures.
  2. Horizontally scalable: each agent can receive a production-like preview instead of competing for a shared staging environment.
  3. Reproducible: the platform can recreate relevant client, network, regional, and runtime conditions.
  4. Push-based: events trigger investigation and follow-up work without polling-dependent human attention.
  5. Atomic: a change is independently testable, releasable, observable, and reversible.
  6. Permissioned: an agent has only the authority needed for its current task and reaches more authority through explicit escalation.
  7. Self-improving: traces, logs, and outcomes become inputs to future agent behavior.

These are requirements for the delivery substrate, not properties that emerge from selecting a more capable model. (Source: sources/2026-08-04-cloudflare-agent-development-lifecycle)

Control-loop shape

Workflow orchestration supplies the durable control plane: record state, retry individual work, wait for a production event or approval, dispatch an agent or test surface, and route the outcome to the next stage. Local-remote parity and ephemeral previews make validation meaningful. Agent-controlled deployment bounds the release loop through flags, cohorts, metrics, and reversibility.

Not a claim of full autonomy

The source presents a design direction, not evidence of a fully autonomous factory operating at a particular safety threshold. It uses a self-driving analogy to argue that a system that works most of the time is insufficient for production autonomy. It does not specify SLOs, permission protocols, model-evaluation methodology, or governance rules for resolving ambiguous customer and product decisions.

Seen in

Last updated · 622 distilled / 1,953 read