Why AI gains are unevenly distributed in your engineering team | Asana’s Arnab Bose
The rise of software factories, the fall of first drafts, and the hidden tax holding back your agents
Why are 75% of knowledge workers using AI, yet only 5% of companies seeing meaningful productivity gains? This week on Dev Interrupted, Asana Chief Product Officer Arnab Bose explains why scaling enterprise AI means shifting from isolated chatbots to fully integrated agentic work management. He breaks down how Asana is turning AI from a tool into a transparent digital teammate with shared memory, full audit trails, and role-based access controls. The conversation closes on Asana's acquisition of Stack AI, the upcoming Command product for R&D teams, and which metrics prove AI ROI.
1. Moving the assembly line to the cloud
Zach Lloyd over at Warp is making a strong case that the software development process is shifting. The industry’s moving away from interactive coding agents run locally by developers and toward automated processes living entirely in the cloud. Having a centralized, shared space where compute is equal across teams is a critical ingredient for scale. He estimates that 20 to 30 percent of issues are fully automatable right now. It fits the mental model perfectly, but since physical manufacturing lines are routinely torn down and retooled for new items, will dark factories need to be gutted and rebuilt from scratch every time new software is shipped? Join our upcoming first-of-its-kind industry roundtable about software factories (featuring Zach!) on August 27th: https://luma.com/88e9rrfg
Read: A guide to cloud software factories for engineering leaders
2. Pushing back against the dark factory
While the cloud factory sounds great, Dev Interrupted alum Dex Horthy is pushing back hard against the concept of a lights-off dark factory. A lot of organizations are trying to achieve massive velocity by cutting out human code reviews and pushing agentic code straight to production based on pass or fail benchmarks. The problem is that a benchmark might tell you if code works, but it tells you absolutely nothing about whether it’s actually maintainable. If you remove human expertise and back pressure, you’re just inviting rot into the software stack. Join our upcoming first-of-its-kind industry roundtable about software factories (also featuring Dexter!) on August 27th: https://luma.com/88e9rrfg
Read: Why Software Factories Fail
3. The danger of bloated agentic org charts
This piece takes a highly critical look at loops, and it’s an absolutely fantastic read. There’s a lot of waste happening as teams delegate processes to background agents, ultimately bloating an artificial org chart. When loops bubble up organically from individual developers rather than a top-down strategy, you end up with a strange variant of Conway’s Law where agents just amplify existing organizational chaos. Not every process scales well with a loop, and sometimes trying to automate a craft actually degrades the final result.
Read: You just hired a million bad employees.
4. The engineering productivity gap: how elite AI teams are pulling away
The engineering productivity gap between elite AI teams and everyone else is real, measurable, and widening every quarter. LinearB recently released its latest benchmark update compiled from 2.7M PRs and 253 orgs. Download the guide to get the complete playbook for scaling your team’s leverage as agents start writing most of the code.
5. Protecting Your First Drafts from AI
Mia Kiraki 🎭 brought a fantastic perspective this week on how AI models handle iterative design and coding. If an agent doesn’t understand the journey and decisions that led to the current input, it’ll almost always make wildly different decisions on how to change it. Passing a project through multiple specialized agents acts like a sander, buffing away all the unique creativity until the output degrades into the most average, normalized version possible. The key to maintaining quality is holding onto precious drafts so the model actually understands the context of what came before it.
Read: An AI that only sees the latest version will never protect your original idea
6. The hidden cost of noisy sub-agents
Rahul Garg from Thoughtworks published some fascinating research on multi-agent coding. He argues that the true cost of a sub-agent isn’t its compute time, but rather how much it pollutes the orchestrator’s context window. Instead of creating sub-agents based on roles like a tester or a front-end developer, they should be organized by “cognitive locality” so they own specific domains of the codebase. If sub-agents constantly interrupt the orchestrator because they lack context, productivity isn’t scaling, it’s just drowning the orchestrator in noise.
Read: The Orchestrator’s Tax
