We Pointed an AI Agent at Our Own Website — A Day of Agentic SDLC

The setup

Most "AI built my website" stories are demos. This isn't one. Before submitting our application to the Anthropic Claude Partner Network, we had a real, unglamorous job: our pre-launch Webflow site carried about sixty catalogued issues — typos, broken components, off-tone stock imagery, missing alt text, a stale list of AI models, and no structured data. So we did the thing we tell clients is now possible: we pointed an autonomous AI agent at it and had the agent do the work, using real tools, on the real site.

Here is what a single working day of agentic software work actually looked like — including where it was genuinely good, and where a human stayed firmly in the loop.

What the agent actually did

The agent didn't "generate a website." It operated our existing stack through tool integrations (MCP): the Webflow Designer and Data APIs, Google Drive, and Notion. In one day it:

• Re-audited all eight pages and produced a versioned audit, reconciled against the previous human audit.
• Swapped three off-tone stock images for authentic photos of our own team.
• Wrote alt text on ~185 images — descriptive for content, deliberately empty for decorative icons.
• Fixed copy and grammar, and removed duplicate form element IDs that were quietly breaking accessibility.
• Added Organization + WebSite structured data site-wide, plus per-page meta titles and descriptions.
• Refreshed our "AI models we work with" lineup — moving Claude to first position and removing deprecated models.
• Published a separate engineering case study (a real upstream open-source contribution).

The measurable outcome: Webflow's built-in accessibility and SEO audit dropped from 52+ flagged issues to low single digits.

Why it was reliable

The interesting story isn't the task list — it's what made an all-day agent run trustworthy.

It verified reality, not its own output. When checking the AI-models lineup, the agent rendered and read each tile rather than assuming; when it couldn't fetch a page normally, it switched methods rather than guessing.

Human review gates on anything public. Content that touched confidentiality was drafted by the agent but published only after a human sign-off. The agent proposed; a person approved.

A continuity system for context limits. Long agent sessions hit memory limits, so the agent kept a two-tier handoff record (local files plus a Notion ops page) — a fresh session could resume exactly where the last left off.

It knew what it couldn't do. Some steps — publishing, creating backups — are deliberately human actions. The agent stopped and asked rather than forcing them.

The takeaway

This is what we mean by AI implementation thinking. The value isn't a model that writes confident text — it's an agent wired into your real tools, checking its work against ground truth, escalating the right decisions to humans, and keeping durable state. Done that way, "AI did it" stops being a slogan and becomes an ordinary, auditable Tuesday.

We ran it on our own site first, on purpose. If it's good enough for our shop window, we can stand behind it for yours.