Dev Interrupted Inference is the new 401k matching and what we’re learning from AI-related outages
22 snips
Mar 13, 2026 They debate the idea of paying engineers with AI compute instead of cash and whether inference costs should fall on employees. They unpack the rise of harness engineering and how planning, docs, and guardrails shape agentic teams. They examine recent AI-related outages and AWS incidents and warn about the pressure to run dozens of autonomous agents. They also share personal agent experiments and laptop-permission risks.
AI Snips
Chapters
Transcript
Episode notes
Inference Could Become A Meaningful Part Of Compensation
- Paying for inference may become a major part of engineering cost and hiring decisions.
- Ben and Andrew note estimates that inference costs could reach ~20% of salary and that token budgets can spike overnight when workflows change.
Roll Inference Costs Into Team Budgets Not Salaries
- Avoid making engineers personally buy inference credits; roll inference into hiring costs instead.
- Andrew argues teams that gain 5x–10x output should be budgeted accordingly so staff aren't forced to pay for tokens themselves.
Harness Engineering Treats Environment As Code
- Harness engineering reframes work: you must build the environment, tools, and guardrails for agents as much as you assign tasks.
- The hosts reference Charlie Gao's playbook and examples like a 3-person team building a million-line product with agentic workflows.