Phase 11ADLC course

Phase 11. Release: the agent ships to production

QA delivered a readiness verdict. Time to ship to production — and simultaneously execute the GTM plan assembled back in phase 3: turning on channels, publishing the landing page, making first touches with the market. Technical release and marketing launch synchronize here.

Deploy is the phase where technical automation has gone furthest of all: CI/CD, IaC, canary, auto-rollback already exist and work. And that's precisely why it exposes the core issue sharpest. If the action has long been automated, why is a human in the release loop at all? This article answers directly: the human stayed in deploy not as the executor of the action, but as the bearer of accountability for risk. ADLC shows this by stripping away the ritual.

The human role today

A release engineer, DevOps, partly SRE: holds the build and deploy pipeline, chooses the rollout strategy (blue-green, canary), manages feature flags, rolls back on problems, makes the "ship or don't" call. Classically this is a release checklist, an on-duty person at the rollout, manual rollback under stress.

Break it down. Build and deploy are machine work already today. The rollout strategy is a choice from a known set under known conditions. Watching rollout metrics and rolling back on regression is an automatable loop. Exactly one human link remains: "ship or don't" — a decision that's actually a signature on risk, not a technical action.

What we hand to the agent

The release agent holds the rollout role. It builds, deploys per the chosen strategy, watches rollout metrics in real time, and rolls back on regression — autonomously, with no one on duty at the button. In parallel, a marketing agent executes the gtm-plan: publishing the landing page, turning on channels, launching first touches in sync with the rollout.

Progressive delivery changes the cost of a mistake. Canary on a small slice of traffic with automatic metric-based rollback makes the rollout reversible by default: if something's wrong, the system reverts before most users notice. And a reversible decision doesn't need a human signature — a signature is needed where rollback is impossible.

Agent architecture

Phase state-machine

Inputs

the release-readiness verdict (phase 10), architecture (phase 5), gtm-plan (phase 3), infrastructure and metrics.

The agent holds the role

Tools: CI/CD pipeline; IaC; feature flags; canary and rollout monitoring; metric-driven auto-rollback; landing-page publishing and channel activation.

Artifact

a production release + activated channels + release notes.

Handoff: live production → Operations (phase 12); market reaction to the launch → Maintenance (phase 13). Notice the fork on reversibility. The release agent classifies the rollout itself: reversible (canary + auto-rollback) it ships autonomously; irreversible (data migration, a public commitment, money movement) it escalates to the principal for a signature. Not "a human presses a button on every release," but "a human is needed exactly where you can't undo it."

Inputs: the release-readiness verdict (phase 10), architecture (phase 5), gtm-plan (phase 3), infrastructure and metrics.
Tools: CI/CD pipeline; IaC; feature flags; canary and rollout monitoring; metric-driven auto-rollback; landing-page publishing and channel activation.
Artifact: a production release + activated channels + release notes.
Trigger: release-readiness = go (phase 10).
Handoff: live production → Operations (phase 12); market reaction to the launch → Maintenance (phase 13).

Notice the fork on reversibility. The release agent classifies the rollout itself: reversible (canary + auto-rollback) it ships autonomously; irreversible (data migration, a public commitment, money movement) it escalates to the principal for a signature. Not "a human presses a button on every release," but "a human is needed exactly where you can't undo it."

Where it breaks

Irreversibility. Data migrations, public commitments, users' money — the cost of a mistake here is asymmetric: a good rollout saves minutes, a bad one costs the business. Appetite for irreversible risk is human, and the release agent has to tell reversible from irreversible apart, not ship everything with equal boldness.

Judgment about timing. "Don't ship on Friday evening," "the client's in peak season right now, let's wait" — business context that isn't in the pipeline metrics. The agent optimizes technical rollout risk; the business context of the moment comes from outside.

Accountability and launch regulation. Payments, personal data, public promises to the market — a launch touches areas where a subject is accountable for consequences. The agent executes the launch; the principal answers for it.

What stays human

The signature on irreversible rollouts and judgment about business timing. And here the course makes an important caveat: this remainder is actively shrinking under progressive delivery itself. The more rollouts get reclassified as reversible (canary, flags, instant rollback), the less often a human signature is needed. The durable core is the irreversible, and it, as always, belongs to the principal (phase 14).

human remainder ≈ 26%

Provocation / thesis

Deploy has long been technically automated; the human stayed in it not as an action but as a signature on risk. ADLC exposes this by stripping the ritual away: once canary and auto-rollback make a rollout reversible, the "release engineer" role collapses — it turns out it was never work, it was readiness to be accountable for the irreversible. Which means a human is needed not on every release, but exactly on the ones you can't take back. Everything else is reversible, and reversible doesn't need a human.

Running case

Vitrina in this phase

The release agent ships Vitrina canary to 5% of stores. The key metric from the gtm-plan — the share of stores with a successful paid transaction — is on target, no payment-loop errors (the very loop that went through the ADR in phase 5, review in phase 9, and regression in phase 10). The agent rolls out to 100% autonomously: the rollout is reversible, no human signature is needed.

In parallel, the marketing agent publishes the landing page and activates local channels from the gtm-plan. The only thing that goes to the founder for a signature is the irreversible: turning on real payment acceptance through the provider (live keys, movement of users' money — can't be undone). Here the release agent doesn't ship on its own, it escalates. The founder signs off — and Vitrina is live in production, taking orders and money.

Notice the breakdown: out of the entire release, a human was needed at exactly one point — authorizing live payments. Rollout, metrics, rollback, landing page, channels — all autonomous. Artifact → Vitrina's production release + activated channels.

In practice

How it actually works — engineering breakdowns

Standalone howto from practice, showing this phase on real code and a working artifact.

Eval as a Release Criterion: Catching Agent Drift Before ProductionA structured eval as the gate before release.
Sending to Kafka Is a Two-Phase CommitDelivery semantics: why sending to Kafka is a two-phase commit, not fire-and-forget.

Building AI-driven delivery in-house?

Designing the ADLC loop: where the agent holds the role and where the human principal remains — for your team and product.

Email me

The transition engine

Next Move Engine — the system that takes a team to an autonomous delivery loop.

Next Move Engine →