Multiplayer AI will revolutionize the way we work | OpenClaw Foundation’s Vincent Koc
Multiplayer AI will revolutionize the way we work | OpenClaw Foundation’s Vincent Koc
The harness pays off for cheaper models, AI took the status updates, and is your favorite old game getting decompiled?
Agents made each engineer faster, but the handoffs between them still run through copy and paste. OpenClaw Foundation chief architect Vincent Koc ran into that friction on his own team, and their answer was to move the engineering work of building OpenClaw inside OpenClaw itself. This week on Dev Interrupted, he joins Andrew to walk through the multiplayer setup behind OpenClaw 2.0, where teammates and their agents can jump into each other’s sessions on shared servers and pick up work when someone goes to sleep. Vincent explains how tools like Git Crawl keep their agents aware of everything happening across the repo, and why the piece that’s still missing is the why behind a decision.
1. The harness pays off for the cheaper model
Giving a coding agent more structure doesn’t automatically pay for itself. Naresh Jain had agents build the same backend-for-frontend service with and without a behavioral harness of specs, examples and deterministic tests, then ran 178 acceptance tests against every result. GPT-6 Sol passed all ten runs either way, so the harness just made it about seven times more expensive. GPT-6 Luna went from 3 of 10 to 10 of 10, and a harnessed Luna attempt still cost about a third less than an unharnessed Sol one. It’s a single greenfield service, but the lesson travels, since model capability decides when the tests come in, not whether you need them.
Read: Are Harnesses Worth the Cost?
2. Glass Town keeps the receipts
Failed runs still eat tokens, which Rizel Scarlett learned after hitting her Codex usage limit while her account automatically topped up. Glass Town, her take on Steve Yegge’s Gas Town, keeps the autonomy of a dark factory but makes every step inspectable afterward. The Entire CLI records each prompt, response and tool call, and links every Git commit to the session that produced it. To rein in costs, she moved the coding roles to a cheaper model on low reasoning, turned off automatic retries, capped each worker at five minutes and replaced the preview LLM with a script. Seeing the work behind a change instead of just its diff is the part I’d borrow first.
Read: Welcome to Glass Town: building a software factory with a paper trail
3. An agent that checks the review queue first
Every agent session is tuned to finish its task and open a PR, which is how a factory manufactures its own bottleneck at review. LinearB’s 2026 Software Engineering Benchmarks Report found agentic AI PRs wait 5.25x longer for pickup than unassisted ones, and only 32.7% of AI-generated PRs merge, against 84.4% of human-written ones. In case you missed it, I posted my fix on the Substack last week, an /sdlc skill that pulls the team’s review pickup time and my waiting PRs, then asks a decision model like Jev whether to clear the backlog before starting anything new. Nothing goes out without my yes or no. I’ll be digging into what this does to the people in the loop at my next software factory roundtable on Oct 29, What your software factory is doing to your engineers, with Marek Poliks and Cate Huston.
Read: Your agents should ask the factory first
Dev Interrupted4. What your software factory is doing to your engineers
Software factories write a growing share of the code your team ships, and the review load lands on the few senior engineers with enough context to catch what’s wrong. In LinearB’s data from more than 8.1 million pull requests, AI-generated PRs merge at 33 percent against 85 percent for human-written ones.
On October 29 at 10am PT I’m hosting Marek Poliks, co-author of The Superdark Factory, and Cate Huston, author of O’Reilly’s The Engineering Leader, for a live workshop on where the software factory is headed. You’ll take away playbooks for measuring review load and team health in your own engineering data using LinearB.
5. AI took the status updates, not the job
Articles calling time on engineering management keep coming, and Liz de Moll, a Senior Software Engineering Manager at Auth0, has lost count of them. She’d happily hand status updates and progress tracking to AI, and argues managers never should have owned that work in the first place. Coaching, prioritizing and keeping teams aligned are another matter, because when teams move faster, the cost of misalignment rises just as quickly. Both times she went on parental leave, her engineers got pulled out of the code and into coordination work that wasn’t theirs. “Removing hierarchy doesn’t remove dependencies,” she writes. “It just removes the people whose job it is to manage them.”
Read: Long live the engineering manager
6. You can reverse engineer and customize basically anything now
Rebuilding a game’s source by hand used to take years, and newer AI models and workflows can now do it in days. The output isn’t as clean, but it works, which is how IsaacRecomp took the 2011 Flash version of The Binding of Isaac to a native 1440p, interpolated to 165 FPS. The catch is that any paid game or app can now be open-sourced against its maker’s will, and there’s no way to prove the code was stolen. Popcar expects some major application to get decompiled and shared, essentially open-sourced by force, and companies won’t take it quietly.
Read: AI is mass decompiling games. That’s both awesome and scary.
7. This week! Come find us at ObservabilityCon and AGNTCon + MCPCon
Most of what we cover week-to-week, from personal agents to planning loops, runs on plumbing that’s shared by everyone. Dev Interrupted is an official media partner of AGNTCon + MCPCon North America, and I’ll be in San Jose on October 22 and 23 to see what comes next in the story we covered today. We’re also on-site today at ObservabilityCon by Grafana capturing the latest from security and observability leadership to keep you ahead of the curve.
BTW in San Jose at MCPCon, more than 3,500 builders will be there, with speakers from Anthropic, OpenAI, O’Reilly, Okta and more digging into MCP and the agent infrastructure we all build on. Come say hi, and use code MEDIA25 for 25% off your ticket.
Subscribe today, no harness required, even on the cheaper models.
Subscribed© 2026 Dev Interrupted
548 Market Street PMB 72296, San Francisco, CA 94104