I built 33 agents in 2025. Here’s what actually worked.
By Alec Newcomb. 7 min read.
Hello, hello! 👋🏻
Everyone’s talking about Claude Code this week. I’ve been quietly using it since August. Here’s what eight months of real usage taught me.
My Claude Code numbers:
- 226 GitHub commits of 299,206 lines of code
- 445,303 lines of content
- 37 custom skills built
- 33 agents deployed
- 4.9 million Claude tokens used
- Longest session: 1 day, 23 minutes
- And that time I deleted 761 files
I’ll share the wins, the disasters, and where I’m going deeper in 2026. Let me be clear: I don’t code. What I do is orchestrate, directing AI to build what I can see but can’t write myself. (Skip the “vibe coding” label. That’s developers being dismissive. This is something new.)
If you want the bigger map of which tool I reach for and when, that lives in my AI tool comparison hub. This post is the behind-the-scenes of building on top of one of them.
Video (MarketingAlec): I Built 33 AI Agents in 2025 (Here’s What Actually Worked)
August: where Claude Code and I started
I was fed up with running prompts in ChatGPT and Claude at the same time. I went looking for a system that would let me be multi-AI from one interface.
Day 1: Tested Claude Code and created a CLAUDE.md file with project instructions.
Day 2: Fixed authentication for my first MCP server (Firecrawl).
Day 3: Set up GitHub (after almost losing everything).
Day 4: Ran my first security assessment and found 19 vulnerabilities.
Early wins that hooked me
- Crawled my entire website and updated my bio automatically
- Set up 9 automated Gmail labels for 90+ newsletters via Google MCP
- Built my first AI agents: a scraping agent and SEO agent v1
- Got multi-model consensus working (GPT, Gemini, and Claude arguing about my code)
First major disaster
I let Claude reorganize my project folder structure in Obsidian.
It worked. And then I spent 3 hours fixing broken wiki-links.
Lesson learned: “Comprehensive” changes need checkpoints.
September to October: building the machine
This is when things got serious. I went from 0 skills to 37 skills for me and my team.
Skills I built:
- Meetings (3-phase automation: raw notes to a standard format, then add people and tasks, with confidence scoring)
- SEO Research (6-phase parallel audit covering Google, Bing, Perplexity, ChatGPT, Claude, and Grok)
- Firecrawl (web scraping workflows)
- Perplexity (real-time research)
- Jina (academic papers, parallel Google search)
The day it passed 41 tests
I built a meeting processing system that automated task extraction, people bios, company details, and priorities.
The moment I watched pytest show 41/41 passing: “Okay, this is actually working.”
Fun discoveries
- Built a skill for creating skills (very meta)
- Discovered AI has strong opinions about folder naming (lowercase only, please)
- Running prompts and skills in parallel changed how fast I move
November to December: the agent era
Here’s the scaffolding I landed on:
August was prompts = detailed instructions.
October was skills = repeatable procedures I invoke.
November was agents = autonomous workers I deploy for a specific problem.
Research was my number one use case, so I started there. The rest flowed from it.
My 33-agent roster by year-end (the top 10):
- topic-ideator
- design-review
- competitor-analyzer
- outline-builder
- seo-orchestrator (the boss)
- research-orchestrator (web research in parallel)
- icp-builder
- gmail-manager (email operations)
- qa-checklist
- documentation
These aren’t simple agents. They’re multi-LLM parallel processing machines doing hours of work at once. A few run in the background, a few run in parallel, the rest only fire when I need them.
The Micro-RACI governance system
After my agents started doing… off-plan things, I built better governance:
- Max 3 iterations before a mandatory handoff
- Evidence requirements (sources plus confidence scores)
- Never auto-publish (always get human approval)
- Log everything, very concisely
- Consistent frontmatter templates over everything
So what? A recent quarterly research project covering more than 270 companies took me two weeks in ChatGPT. I finished it in six hours over two days with Claude Code. My wife is very happy, because even she (my chief AI skeptic) is seeing the progress.
The numbers
Productivity wins
- MCP bloat: 85% token reduction (77K down to 8.7K)
- Meeting processing: 50 meetings a month, turned into people and tasks in minutes
- Multiple LLMs in parallel, easily (Gemini, Claude, ChatGPT, Grok, DeepSeek)
Money saved and made
- Replaced 5 separate SaaS tools (around $600 a month in savings)
- Competitive intelligence: extracted 314+ Meta ads in one session
- Outreach: scored and prioritized 200+ prospects automatically in a day
What I taught Claude Code
- How “Alec” likes to work (we’re partners)
- Short, well-defined processes with gates like Micro-RACI are the way to go
- Document and back up as you build (outdated docs are a stop-the-line issue)
What Claude taught me
- Doing it right beats doing it fast. Never skip steps.
- Tedious, systematic work is the way (I have workflows that run for 2 to 3 hours)
- A single source of truth prevents chaos
- Background processes need gates and kill switches
The real moments
- Claude is wrong sometimes. It happened enough that I built Micro-RACI around it. (Public benchmarks still put model hallucination rates in the high single digits to low double digits, depending on the task, so I treat every output as something to verify, not trust.)
- The day I accidentally deleted 761 files and GitHub helped me restore them
- The day I made a mess of a GitHub merge. Took me two days to fix. Not pretty.
- Agents reaching for the wrong models (they love defaulting to older, cheaper ones instead of the current frontier release I actually want)
What’s next
- Model-first reasoning rollout (to reduce hallucinations)
- 6 agents queued for v2 and v3 upgrades (better handoffs, sharper workflows)
- Explicit constraint modeling
- State tracking built into workflows
- More background agents
- Overnight agents
My 2026 philosophy
This year taught me that AI isn’t about replacement. It’s about collaboration. In parallel. At a scale that still boggles my mind.
My best work happened when my core trio (Claude, Gemini, ChatGPT) pushed back on bad ideas, fixed broken processes, and argued with each other about what to do.
The future isn’t AI doing your job. It’s AI arguing with you about the best way to do your job, and then actually listening. If you’re still picking one assistant and staying loyal, that’s the part I’d rethink first. The marketers pulling ahead are routing each job to the right tool, which is the whole idea behind my AI marketing approach.
Cheers, Alec
P.S. This started as a test. If you want more posts like this, reply with a hell yes.
P.P.S. If you read this and felt left out, confused, or late to the game, you’re not alone. Watch this video. We’re all feeling it.
Want the build log behind the agents?
Twice a week I write up what I’m actually wiring together with Claude Code: the agents that earn their keep, the merges that blew up, the governance I bolt on after something goes sideways. No polished case studies, just the real workbench. If you want the next experiment before it’s tidy, subscribe free.