The Productivity Paradox: Code Is Written, Releases Stall
The emergence of autonomous coding agents has created the illusion of the imminent automation of the entire software development industry. However, large-scale measurements reveal an unexpected gap between code generation and the delivery of business value.
According to a study by the National Bureau of Economic Research (NBER), which analyzed telemetry from over 100,000 engineers on GitHub, using AI agents increases commit-level activity by 180%. Yet, the increase in completed projects is only 50%, and no more than 30% of the generated volume reaches final release. This creates a classic "weakest link" effect: making one stage drastically cheaper instantly overwhelms adjacent stages in the delivery pipeline.
Where the Saved Time Is Lost
METR research confirms the mixed nature of this effect: depending on the context and codebase complexity, developer productivity with AI fluctuates from a 38% drop to an 18% increase. The primary obstacles emerge at the intersection of writing code and deploying it:
- The Code Review Crisis. At Linear, following agent adoption, the volume of completed tasks and merged code increased by 50%, yet review became a critical bottleneck: engineers face overload and the temptation to approve pull requests with a boilerplate
looks goodwithout thorough inspection. - Rising Instability. According to DORA (DevOps Research and Assessment) reports, accelerated code generation is accompanied by higher auditing overhead and increased release instability.
- Organizational Barriers. Microsoft's Work Trend Index research shows that company processes and culture have twice the impact on AI ROI compared to individual developer skills (67% vs. 32%), yet only 19% of professionals work in teams with processes ready for this shift.
Why Traditional Issue Trackers Are Losing Relevance
Traditional tools like Jira or Linear capture static task statuses but fail to reflect dynamic changes. When business requirements shift in chats and neighboring teams update APIs in parallel, a card marked In Progress or Done turns into outdated cache.
If an agent successfully completes a task based on the initial specification while project context has already shifted, the output loses its value. Human managers are forced to manually correlate dozens of disconnected events, becoming the system's bottleneck.
A Paradigm Shift: From Manual Control to Process Automation
The solution lies not in attempting to build a single superintelligent AI manager (which leads to unchecked agent self-attestation), but in restructuring the operational model of software development:
The human's role is to make architectural and product decisions, define goals, and sign off on risks. The management platform's role is to continuously track contracts, dependencies, and readiness verification across stages.
The key metric for AI adoption success becomes not the number of tokens spent or commits generated, but the share of code that reaches production without excessive manual coordination overhead and rework.