Case Study6 min read

One Agent, an Assembly Line, or Agent + Editor? We Built the Same AI Agent 3 Ways

Does architecture matter more than model choice? We built the same Reddit-posting agent three different ways, ran 8 model mixes through a 14-criteria AI judge panel, and the most expensive configuration lost twice. Here is the full scoreboard.

By Danny Lev, Founder & CEO

Everyone argues about which model to use. Almost nobody tests which architecture to use. So we built the same Reddit-posting agent three different ways, ran 8 model mixes across them, and graded every run with a 14-criteria AI judge panel. The most expensive configuration lost. Twice.

The three architectures

Same job in all three: read a set of source pages, write 4 publish-ready Reddit posts with a target subreddit and a real image, and obey each community's rules. What changed is how the work is split:

1 · One mega-agent. A single agentic prompt does everything: research, writing, images, compliance. Simplest to build, hardest to control.
2 · Assembly line. Four exact single-task steps: plan → write → attach images → final compliance pass. No open-ended loops; each step does one thing.
3 · Agent + editor. A writer drafts everything, then a second model edits: compliance, voice, images. The architecture our Reddit Post Agent template ships.

The setup

Every configuration ran on the same 3 dataset samples (each a set of live source pages), and every post was graded by the same judge panel: Claude Opus 4.8 scoring 14 criteria, 0–1 each, with browsing enabled so it verifies claims against the actual source pages. Criteria include content value, community compliance, human voice, engagement potential, originality, and whether a real image was attached. All of it ran as Evaligo experiments with a cost cap.

All 8 configurations, ranked

Setup Judge score Human voice $ / post
Assembly line · Opus writes, mini rest
★ best score per dollar
85.5% 0.73 $0.04
Agent + editor · mini writes, Opus edits 85.4% 0.77 $0.23
Assembly line · Sonnet writes, mini rest 85.2% 0.67 $0.04
Agent + editor · Opus writes, mini edits
★ best human voice of any strong setup · what our template ships
84.6% 0.88 $0.10
Assembly line · Opus everywhere
priciest of its group — 3rd of 4
84.1% 0.50 $0.24
Agent + editor · Opus writes, Opus edits
priciest of its group — 3rd of 4
83.0% 0.52 $0.49
Assembly line · mini everywhere 82.3% 0.60 $0.02
Agent + editor · mini writes, mini edits
cheapest — and the most human voice of all
80.9% 0.90 $0.01

3 samples per configuration, 4 posts per run. “$ / post” is the flow cost only. Judge score = the average of all 14 criteria.

5 things we learned

1

“Best model everywhere” lost twice.
Opus-on-every-step finished 3rd of 4 in both architectures, at up to 13× the winner's cost.

2

Expensive setups sound the least human.
All-premium configs scored 0.50–0.52 on the human-voice judge; the cheapest scored 0.90. Over-polish reads as AI.

3

Routing beats brute force.
The strong model only on the step that moves quality: higher score than all-premium at 84% lower cost, about $0.04 per publish-ready post.

4

Sometimes the prompt is the bottleneck, not the model.
The assembly line's human-voice score was stuck near 0.40 through several model upgrades. What fixed it was a prompt change: hard 150-word cap, topic in the first sentence. Same models, +45 points on voice.

5

The cheap floor is surprisingly high.
Mini-everywhere landed at 81–82% for $0.01–0.02 a post: fine for volume drafts, wrong for the posts that represent you.

Bottom line: pick the architecture first, then route models per step. We ship the agent + editor setup with a strong writer and a cheap editor — statistically tied with the top score and the best human voice of any strong config. It's a ready-made template you can run in minutes.

Method notes, honestly

Three samples per configuration is enough to expose the big gaps (single runs wobbled a few points, which is exactly why we averaged), but the 1-point differences at the top of the table are ties, not rankings. The judge is a single model (Opus 4.8) applying written criteria; a different judge would shift absolute numbers, though in our earlier checks the ranking held. Every experiment, judge goal, and cost figure in this post came straight from the Evaligo experiment reports — the same feature you can point at your own flow.

#case study#agentic workflows#architecture#model comparison#llm judge
DL

Danny Lev

Founder & CEO at Evaligo

Founder of Evaligo. Building AI automation tools that help teams ship faster. Previously led engineering at enterprise AI companies.

10+ years in AI/ML engineeringBuilt systems processing millions of AI requests

Ready to Build This?

Start building AI workflows with Evaligo's visual builder. No coding required.

✓ No credit card✓ Free tier available✓ Deploy in minutes

Need Help With Your Use Case?

Every business is different. Tell us about your specific requirements and we'll help you build the perfect workflow.

Get Help Setting This Up

Free consultation • We'll review your use case • Personalized recommendations