DEV Community: Orbyt Labs

One developer, 298 agents in a single workflow, and the bottleneck that moved

Justin Bartak — Fri, 30 Oct 2026 13:00:15 +0000

One month of my transcripts records 2,800 agent runs, and the largest single workflow held 298 agents, running eight at a time.

The removal test: pull the model out of your product. If what's left still works, you built a tool with an AI feature. If it collapses, you built something AI-native.
I built Orbyt solo in 32 days for $400 in model spend: 243,000 lines and 4,124 tests at launch. It's past 720K lines now, with 50,809 tests.
Most teams point agents at tasks and then watch every keystroke, which is just slow typing with extra steps. Then the catch: an agent aimed at the wrong feature builds it fast and well and never asks whether it should exist.

Read the full article: Build From the Model Out.

This summary was written with AI assistance (Claude) from the full article at https://www.orbytlabs.ai/blog/ai-native-frontier-building-core-concepts.

My agent team found 16 accepted findings to one agent's 9 and resolved zero briefs

Justin Bartak — Thu, 29 Oct 2026 13:00:12 +0000

On six competitor briefs, a research team produced 16 accepted findings against a single agent's 9, resolved none of the briefs, and cost $16.98 against $7.84.

The single agent resolved one brief and the team resolved none, so I did not keep the team. That was the rule I declared before the run. More accepted findings did not add up to more accepted work.
I've already watched internal checks agree and still miss the task. My page generator and sitemap read from the same incomplete list, so they matched, every check passed, and the requested pages returned 404.
In a synthetic supplier test, an answer can cite every document and still fail. "Alder is cheaper, so choose Alder" saves $1,000 and ignores that Alder deletes backups in 60 days when the buyer allows 30.

Read the full article: The Agent Team That Found More and Failed

This summary was written with AI assistance (Claude) from the full article at https://www.orbytlabs.ai/blog/which-agent-architecture-fits.

An AI writes the most common version of your brand, and you never see the draft

Justin Bartak — Thu, 29 Oct 2026 06:27:28 +0000

Brand mentions correlate with AI visibility roughly three times more strongly than backlinks, which puts the biggest brand lever off your site entirely.

An engine that meets two versions of the same number trusts neither, so I keep every citable figure in one tested source of truth and the build fails when any surface drifts.
The machine needs a who it can verify. My identity resolves through Wikidata and ORCID, and a 15 million data point dataset across 3,445 roles and 81 cities sits under a DOI so figures trace to their origin.
I measure the description engines give of my own work. The published probe reads zero, which is what a starting line looks like when you count instead of assume.

Read the full article: Described More Than Seen.

This summary was written with AI assistance (Claude) from the full article at https://www.orbytlabs.ai/blog/brand-in-the-answer-economy.