Why Your Company’s “AI‑First” Strategy Might Not Be Real AI‑First

AI Assist vs AI‑First are Different

Peter Pang distinguishes "AI assist"—where engineers use tools like Cursor or ChatGPT to speed up existing tasks by 10‑20%—from "AI‑first," which redesigns processes, architecture, and organization so that AI is the primary builder and humans provide direction and judgment.

Three Bottlenecks That Kill You

Peter identifies three bottlenecks:

Product management bottleneck : Traditional PM cycles take weeks, but an AI Agent can deliver a feature in two hours, making long planning cycles obsolete. PMs must become "product‑thinking architects" to keep up.

QA bottleneck : Agents deliver features in two hours while manual QA takes three days. Replacing human QA with an AI‑driven testing platform aligns verification speed with construction speed.

Labor bottleneck : Competitors use 100× more people; CREAO achieves comparable output with 25 engineers by redesigning the workflow.

All three stages—design, implementation, testing—must be AI‑native; any manual step stalls the pipeline.

Fix the Codebase First

Before building the new system, Peter consolidated fragmented repositories into a single monorepo so the AI Agent can see the entire codebase, enabling it to reason about cross‑service impacts and run integration tests.

This "harness engineering" principle states that the more of the system an Agent can inspect and modify, the greater the leverage.

Technology Stack

Infrastructure: AWS – auto‑scaling, circuit‑breaker rollbacks, CloudWatch as a centralized observability hub with structured logs and 25+ alerts.

CI/CD: GitHub Actions – a deterministic six‑stage pipeline (validation, build, test, deploy, production test, release) that enforces type checks, linting, unit/integration tests, Docker builds, Playwright E2E tests, and environment consistency without manual overrides.

AI Code Review: Claude Opus 4.6 – three parallel reviews per PR (code quality, security scan, dependency scan) that act as mandatory gates, not suggestions.

Self‑Healing Loop

Every day at 09:00 UTC, an automated health workflow runs: Claude Sonnet 4.6 queries CloudWatch, analyses error patterns, and posts a management health report to Microsoft Teams.

An hour later, a triage engine clusters CloudWatch and Sentry errors, scores them on nine severity dimensions, and creates Linear tickets with sample logs, affected users, endpoints, and suggested investigation paths.

The system de‑duplicates tickets, reopens recurring issues, and automatically closes tickets once the CI pipeline validates the fix.

Toolchain and Full Workflow

Features are gated by Statsig feature flags with gray‑scale rollout and instant kill capability. Graphite manages PR rebasing and merge queues. Sentry provides structured exceptions that feed into the triage engine alongside CloudWatch. Linear surfaces AI‑generated tickets to engineers.

New feature path (eight steps): architect defines task → Agent decomposes, plans, codes, tests → PR opens, three‑round Claude review → full CI validation → Graphite merge queue → six‑stage deployment → feature‑flag gray rollout → automatic rollback on degradation.

Bug‑fix path (five steps): detection → Claude triage creates ticket → engineer investigates (AI‑generated diagnosis) → same review, CI, deployment pipeline → triage engine verifies resolution and closes ticket.

14‑Day Data

In two weeks CREAO averaged 3‑8 production deployments per day, whereas the previous model could not even release once in that period. Features and bugs are shipped the same day they are conceived, with real‑time A/B validation. User engagement and paid conversion rose because the feedback loop tightened.

Future Engineer Roles

Architect : 1‑2 people design SOPs for AI, build testing and triage infrastructure, define system boundaries, and decide what “good” means. They must critically evaluate AI output, spot missing failure modes, and guard against technical debt.

Operator : The rest of the team executes AI‑assigned tasks, investigates tickets, validates fixes, and performs UI or CSS tweaks. The work shifts from raw coding to oversight and decision‑making.

Who Adapts Fast

Junior engineers adapt quicker because they feel empowered by AI‑amplified impact, while senior engineers struggle as AI compresses two‑month workloads into an hour, challenging long‑standing habits.

Management Collapse

Peter’s management time dropped from 60% to under 10% as AI handled alignment, meetings, and feedback. He now spends most of his day coding, designing SOPs, and maintaining the harness system, leading to a more relaxed team culture despite higher workload.

Beyond Engineering

Other functions—product, marketing, growth—also run on AI‑native workflows: changelogs, video demos, social posts, and health reports are generated automatically. Any function that remains manual becomes the new bottleneck.

Advice for Different Audiences

Engineers : Shift value from raw code output to decision quality, critique AI suggestions, and develop product‑sense.

CTOs & Founders : Start by shortening PM cycles, build a test‑harness system before scaling Agents, and begin with a single architect who proves the concept.

Industry : Major AI labs converge on structured context, specialized agents, persistent memory, and execution loops; model upgrades (e.g., Claude Opus 4.5 → 4.6) drive rapid capability gains, suggesting a future where one architect with an Agent can replace dozens of employees.

CREAO’s AI‑native engineering architecture (source: @intuitiveml)

Logic for breaking the three bottlenecks (source: @intuitiveml)

Daily health check and triage engine (source: @intuitiveml)

Statsig, Graphite, Sentry, Linear collaboration (source: @intuitiveml)

Unified pipeline from feature development to bug fix (source: @intuitiveml)