The math on agent labor just shifted from theoretical to undeniable.
The Summary
- Asana used OpenAI Codex to replace their outdated testing infrastructure in two weeks, completing work their team estimated would take five years of human engineering time
- Total cost: roughly $12,000 in API calls and compute
- This isn't a proof of concept anymore. This is production engineering work at 130x speed compression and a cost structure that makes traditional dev planning obsolete
The Signal
Asana's engineering team had a problem every software company knows: legacy test infrastructure slowly rotting while the roadmap screams for new features. The work to modernize it sat in the backlog with a five-year timeline. Too big to tackle, too boring to prioritize, too risky to ignore forever.
They pointed Codex at it. Two weeks later, done. For twelve thousand dollars.
"Five years of planned engineering work collapsed into two weeks for the cost of a mid-level developer's monthly salary."
The implications split into three layers. First, the obvious one: agent labor is here and it's absurdly cheap compared to human time. $12K would barely cover benefits for one engineer for a month. Asana just bought 130 weeks of equivalent output for that price. The unit economics don't just favor automation, they make the traditional hiring model look like a rounding error.
Second layer: this wasn't greenfield work or a side project. This was core infrastructure replacement. The kind of work that requires understanding existing systems, making architectural decisions, and touching production code without breaking things. Codex handled it. That suggests the threshold for "work agents can do" just moved from peripheral tasks to load-bearing systems.
Key shifts this represents:
- Backlog prioritization changes when timeline compression is this extreme
- Technical debt becomes a solvable problem with different resource constraints
- The planning assumption that "we can't afford to modernize legacy systems" no longer holds
Third layer, the one that matters most: Asana didn't build this capability in-house. They used an API. No specialized AI team, no six-month integration, no custom training pipeline. Point, click, pay per token, ship. That's the Web4 pattern. Companies don't need to become AI companies to use AI labor. They just need to know which agents to hire.
The Implication
If you're running engineering, your backlog just became radically more fungible. Work that you've been deferring because "we don't have the people" might cost less than your monthly AWS bill. Start inventorying tasks by agent-readiness, not just business priority.
If you're an engineer, this isn't the "AI will replace you" story. It's the "your leverage just went vertical" story. The engineers who pointed Codex at this problem still made the decisions. They just stopped being the bottleneck for execution. Learn to direct agents, and your output ceiling disappears.