Governing AI deployments
with approvals and rollback.

Overview
Enterprises were shipping AI agents to production with no versioning, approvals, or rollback. I conceptualized and designed the Agentic Development Life Cycle (ADLC) to bring software engineering discipline (staging, PR reviews, version history) to agent deployment on StackAI.
Company
StackAI
Feature
ADLC deployment governance
Timeline
Q1–Q2 2026
Impact
~250 users followed the approval-gated publishing flow, and environment keys went from zero to 644 in 3 months.

PROCESS

Personas and mental models Benchmark, GitBook and other product inspiration Iteration 1, first design iteration Iteration 2, second design iteration Iteration 3, third design iteration

1. "We need to build this into the product."

An enterprise customer was governing AI agent deployments by hand, a 6-phase process their CISO built outside StackAI. After interviewing the team and mapping how they worked, we saw the potential to productize it. He called it the Agentic Development Life Cycle. We co-designed with enterprise security and platform teams so the product reflected how they actually govern deployments.
Personas and mental models

2. Benchmark

Drew from GitBook (not GitHub), aimed for granularity over binary on/off, fixed the naming inconsistency (product said "SDLC", marketing said "ADLC").
Benchmark, GitBook and other product inspiration

3. Iteration 1

First design iteration, exploring layout, stage granularity, and how ADLC settings map to the existing project model.
Iteration 1, first design iteration

4. Iteration 2

Second iteration, refining environment stages, diff views, and the convert-to-ADLC onboarding flow.
Iteration 2, second design iteration

5. Iteration 3

Third iteration, final direction for per-project settings, approval workflow, and environment variable handling.
Iteration 3, third design iteration

SOLUTION

Submit a Pull Request for Review. Enterprises were shipping agents with no staging gate, so changes jumped straight into production. Modeling the CISO's ADLC process, editors request admin approval before promoting to Staging, Development, or Production, bringing software-style review into agent deployment instead of trusting ad-hoc process outside the product.

Review Requests & Accept Changes. Security and platform teams needed to inspect diffs before anything reached production, not just flip a binary on/off switch. Admins review each request, see what changed, and approve into Production so governance stays granular and auditable where agent risk actually lives.

Review changes dialog with node diffs before deploying to staging
Active deployments and deployment history across Development, Staging, and Production

Per-Stage Environment Variables. Hardcoded keys and URLs made the same agent unsafe to promote across environments. Variables live outside the workflow graph so Development, Staging, and Production can share one agent definition while secrets and service endpoints stay stage-specific, matching how teams already govern other software.

Environment variables managed across Development, Staging, and Production

Share Agents via Agent Grid. A governed agent only creates value when the org can find and use it. Once approved, builders publish finished agents to the Agent Grid so deployment discipline connects to day-to-day discovery instead of leaving production agents stranded in private projects.

INSIGHTS

Enterprise sales requirement

Became a standard requirement in every enterprise sales conversation.

Validated before it was built

Concept validated externally before it was built, a customer coined it first.

A universal governance gap

The "I need governance now, but I built this without it" case turned out to be universal.

OTHER PROJECTS