Grinberg & Gill (Claude channel) — Factory and the dark factory
▶ Watch on YouTube · Claude · 10:38
Factory is bringing autonomy to software engineering with Droids, AI coding agents that handle enterprise work. In this conversation, Co-founder and CEO Matan Grinberg sits down with Anthropic’s Kai Gill to talk about his journey founding Factory, how Claude’s frontier models handle long-horizon tasks across large codebases so teams can trust agent plans with less oversight, and what it will take to build a fully autonomous “dark factory” for software.
TL;DR
A 10-minute customer interview on the Anthropic Claude channel. Matan Grinberg, co-founder and CEO of Factory, who left a physics PhD on string theory in 2023, talks with Anthropic’s Kai Gill. Factory’s agents are called Droids. Four claims:
- Reduced oversight is the unlock. The biggest gain from Opus is “reducing human oversight.” Plans used to be closely monitored; now “you literally might not even look at them,” and going back to one is “a very crazy corner case.” A bad plan may only show up “billions of tokens down the line.”
- The dark factory. Droids “lights off 24 hours a day, 7 days a week” would monitor a deployment, see a fall in a completion rate, test in a sandbox, propose a fix, ping a person, deploy and monitor again. “We’re not fully there yet.”
- A lesson-carrying migration. At one bank, a Factory engineer sat beside the bank’s engineers for a first COBOL modernisation while the Droid extracted lessons. On the next one, “the amount of human involvement was 90% less.”
- The harness has to keep being replaced. “A model today could do what a model plus a harness could do 2 years ago.” As models improve, “if you are lazy, and you don’t change your harness, it’s going to be superseded.” Hence “in one hand a sledgehammer” and a monthly rethink.
Key claims
Why industry, and being early
Code generation is most interesting in “the world’s largest enterprises, some of the messiest codebases where some of the code is 30 years old,” which is why he left academia. Gill recalls him saying that being too early is the same as being wrong. Grinberg calls 2023–25 a blessing in disguise, since he knew nothing about running a company and “would have missed the opportunity.”
Trust as behaviour change
“As you get used to delegating tasks to agents that can go for days or weeks on end,” people get used to not touching something for days. Seeing meaningful output builds trust. He also says the product must “respect their intelligence as developers” and be configurable “to the silhouette of their organization.”
Which model where
“If you have the CTO of a company implement every line of code, that is not an efficient use of their time.” Grinberg uses Opus for planning and for orchestrating many moving parts.
Three hundred things
“Every company wants to ship 30 things, but they can only ship three and a quarter … Well, now you can ship 300 things. But will those 300 things actually make a difference?” The human question becomes “what really is our core competency?”
What blocks it
“Three concentric circles. At the core, there’s the model, then there’s the harness, and then there’s the company and the workflows of the company as a whole.” The interview does not say what in the outer circle is blocking; the answer stays on the harness and on the partnership with Anthropic.
Dynamic-capabilities reading
digital-seizing/strategic-agility: rethinking the harness monthly, with “a sledgehammer to whatever we thought was the best solution.”strategic-renewal/collaborative-approach: engineers delegating multi-day work, trust built through output, and a Factory engineer beside a bank’s engineers on the first pass.
How it connects
- agent-oversight-and-delegation — plans no longer read, and trust as behaviour change.
- agent-harness — the model absorbing the harness, in a harness vendor’s words.
- agent-fleet-management — the dark factory as an always-on monitoring loop.
- agentic-engineering — a legacy migration with lessons carried forward.
Related in this wiki
- Lloyd (Warp) — another vendor’s software factory, with the loop defined in code.
- Kilpatrick (Sequoia) — “the model eats the harness,” and the lock-in objection.
- Chase (Interrupt London) — model, harness and context as what improves.
- Hack (ClickUp) — another former coding-agent founder on building through the same model generations.
What was actually ingested
The full interview (10:38, length_seconds: 638) from the human-written caption track via the yt-dlp fallback, 64 lines. The edit cuts between speakers and the track has no speaker labels; attributions follow the sense and the description. No stills: a frame probe showed two people talking.
Linked entities and concepts
- Anthropic — publisher; Claude is the supplier in the story.
- agent-oversight-and-delegation, agent-harness, agent-fleet-management, agentic-engineering — updated in this ingest.
- dynamic-capabilities — the two cells tagged above.
Dangling (single-source mention, deferred): Factory, Matan Grinberg, Kai Gill.
Scope and reliability
A customer story published by the supplier, with a link to a marketing write-up in the description. The bank is unnamed, and the 90% is one before-and-after comparison reported by the vendor, on a second migration that also benefited from the first having been done. Whether plans should go unread is not examined. The interview is useful as a plain statement of what a coding-agent vendor now sells: less human oversight.