I Ran 6 AI Agent Workflows for 90 Days — Only 3 Survived
· Insights
Ninety days of running agentic coding workflows in real projects. Three earned a permanent slot, three got deleted. Here is the honest scorecard.
Last updated: August 3, 2026 · 8-minute read
I sell software by day and build it at night, which means my tolerance for a workflow that needs babysitting is roughly zero. Over the last 90 days I ran six agentic coding workflows across real, shipping projects — this portfolio, TalkDrive, a Tauri desktop app, and a pile of Hugging Face Spaces.
Three of them are still running today. Three are deleted. This is the scorecard, with the numbers I actually tracked: time saved per week, how often I had to undo the agent's work, and whether it survived a bad week.
How I scored them
Every workflow got the same three questions after 90 days:
That last one killed more workflows than the first two combined.
Survivor 1 — The scoped refactor agent
What it does: one agent, one file tree, one instruction — "migrate every component in src/components/site to the new token system, do not touch anything else."
Why it survived: the blast radius is defined before the agent starts. It reads a CONVENTIONS.md, it works inside a git branch, and every run ends with npm run test. Revert rate: 8%. Roughly 3 hours saved per week during a migration week, near zero otherwise.
The key detail people skip: I give the agent the diff shape I expect, not just the goal.
## Allowed edits - src/components/site//.tsx (className strings only)
## Forbidden - Any change to props, hooks, state, or exports - Any new dependency
## Definition of done - rg "text-white|bg-black|#[0-9a-f]{6}" src/components/site returns nothing - npm run test passes
Constraints beat prompts. Every time.
Survivor 2 — The content pipeline
Every post on this site goes through a scheduled GitHub Action that regenerates the sitemap and RSS feed, builds web stories, cross-posts with a canonical URL, and pings the indexing APIs. I wrote about the whole thing in the automated blog pipeline post.
Why it survived: it's deterministic. The agent writes the draft; a plain Node script does the publishing. The moment I tried to let the agent also handle publishing, it started inventing slugs. Splitting "creative" from "mechanical" is the whole trick.
Time saved: ~4 hours/week. Revert rate: effectively 0%, because a broken script fails loudly in CI instead of silently in production.
Survivor 3 — The review-before-review agent
A pre-commit agent that reads the diff and answers three questions: what changed, what could break, what did you not test? It doesn't approve or block anything. It writes a paragraph.
Why it survived: it's advisory. It costs nothing to ignore, and about once a fortnight it catches something genuinely embarrassing — a missing null guard, an env var only defined locally, a console.log with a token in it. Related reading: free AI code review in GitHub Actions.
Casualty 1 — The autonomous multi-agent "team"
Planner, coder, reviewer, tester. It demos beautifully. In practice, each handoff loses context, and by the fourth hop the tester is validating a plan the coder already abandoned.
I got value out of exactly one variant of this — the one described in Agency-Agents, where agents share one filesystem and one task file instead of passing messages. Message-passing teams: deleted. Shared-artifact teams: kept, but under Survivor 1's rules.
Casualty 2 — The always-on repo watcher
An agent watching for TODOs and opening PRs. Week one: delightful. Week three: 14 open PRs, none of which I wanted, all of which needed a review. The cost of an agent is not the tokens — it's the queue of decisions it hands you.
Casualty 3 — The "just ask it anything about the codebase" chat
Killed not because it was bad, but because a well-configured CLAUDE.md plus a memory MCP server made it redundant. See the memory + MCP setup — once the agent had persistent project context, the separate Q&A bot had no reason to exist.
The three rules I'd give my past self
1. Scope beats intelligence. A mediocre model with a tight boundary outperforms a great model with a vague one. 2. Never let an agent own the irreversible step. Agents draft; scripts deploy. 3. Count the decisions it creates, not just the ones it removes. A workflow that hands you 14 PRs has negative value.
What I'm testing next
Agent skills as a distribution format — packaging a workflow as a reusable skill instead of a prompt I paste. Early results are promising enough that it's the subject of Lord of the Skills.
If you're building on this, start with one workflow, define the forbidden list before the goal, and give it 30 days before you judge it. Most agentic setups fail in week three, not week one.
More in the AI Agents topic hub, or see what I'm shipping on the projects page.