Agentic Coding Metrics and the 2026 CFO Shift
Jellyfish data reveals a 2x increase in merged pull requests but highlights the 'agentic barrier' of human attention limits. Engineering leaders must shift from output metrics to business outcomes to satisfy CFO scrutiny in 2026.
The Shift from Experimentation to ROI
The landscape of AI-assisted development has transitioned from speculative experimentation to measurable operational impact. Data from Jellyfish, tracking approximately 250,000 developers and 40 million pull requests, confirms that AI tools have doubled the rate of merged pull requests. This 2x increase in raw code throughput validates the utility of agentic coding tools, moving the conversation from "if" to "how" for enterprise adoption. However, this productivity gain is not uniform; it is heavily skewed toward newer codebases and AI-friendly languages like Python and TypeScript, while legacy, distributed systems show minimal improvement due to complex context requirements.
The Agentic Barrier and Workflow Friction
A critical bottleneck, termed the "agentic barrier," limits human scalability. Most developers effectively manage only one to two agents, with experienced engineers capping out at four concurrent tasks due to attention limits. To break this ceiling, organizations must move from interactive supervision to fully autonomous handoffs. Furthermore, the quality of agentic output requires scrutiny: AI-generated pull requests have a 60% merge rate compared to 80% for human-written code. This 40% failure rate suggests that current workflows are not optimized for agent output, leading to significant waste and the need for new selection and review mechanisms.
2026: The Year of the CFO
As token costs skyrocket due to more complex models and longer agent turns, the focus shifts to financial accountability. Engineering leaders can no longer justify spend based on adoption rates or raw activity. The "Year of the CFO" demands a shift from output metrics (pull requests, tokens) to business outcomes (revenue, feature velocity, market feedback). Companies must demonstrate that increased coding capacity translates to shipped value, not just more code. This requires rethinking the entire Software Development Life Cycle (SDLC), as traditional metrics fail to capture the true economic impact of AI acceleration. Success now depends on aligning engineering velocity with go-to-market capabilities and business strategy.
Strategic Implications
Organizations must invest in dedicated AI enablement teams to bridge the gap between tool availability and effective adoption. The era of "vibe coding" is giving way to structured, autonomous workflows that require robust context engineering and clear architectural boundaries. Leaders must focus on continuous learning and organizational evolution to maintain a competitive edge in an accelerating technological landscape.
Key insights
-
AI tools have doubled the number of merged pull requests, confirming significant productivity gains in code throughput. This 2x increase is consistent across the industry, validating the core utility of agentic coding assistants.
Impact: Engineering teams can leverage this throughput to accelerate feature delivery, provided they address downstream bottlenecks in review and deployment.
-
Human attention limits create an "agentic barrier" at four concurrent agents. To scale beyond this, developers must shift from interactive supervision to fully autonomous agent workflows.
Impact: Organizations must redesign development processes to support autonomous handoffs, reducing human-in-the-loop friction and enabling higher concurrency.
-
Agentic pull requests have a 60% merge rate, significantly lower than the 80% rate for human-written code. This indicates a high failure rate and waste in current agentic workflows.
Impact: Companies need to implement better selection and review mechanisms for AI-generated code to reduce waste and improve overall efficiency.
-
Token costs are rising sharply as models become more complex and agents perform more turns. This financial pressure is shifting the focus from adoption metrics to ROI and business outcomes.
Impact: Engineering leaders must demonstrate clear financial returns on AI investment to secure continued budget support from CFOs and executives.
-
AI productivity gains are uneven, with newer codebases and AI-friendly languages showing significant improvements while legacy systems lag. This disparity highlights the need for context engineering and codebase modernization.
Impact: Organizations should prioritize modernizing legacy code or implementing robust context management to unlock AI benefits across their entire software portfolio.
Action items
-
Shift performance metrics from raw output (pull requests, tokens) to business outcomes (revenue, feature velocity). Align engineering goals with broader business objectives to justify AI spend.
Impact: This alignment ensures that AI investment directly contributes to company growth, satisfying CFO scrutiny and securing future budgets.
-
Invest in dedicated AI enablement teams to drive adoption and best practices. These teams should focus on context engineering, training, and tool selection to overcome the "agentic barrier."
Impact: Dedicated resources accelerate organizational learning and ensure consistent, effective use of AI tools across all engineering teams.
-
Redesign workflows to support autonomous agent handoffs. Move away from interactive supervision for routine tasks, allowing agents to work independently with clear specifications.
Impact: Autonomous workflows break the human attention limit, enabling developers to manage more concurrent tasks and increase overall throughput.
-
Implement robust review and selection mechanisms for AI-generated code. Address the high failure rate of agentic pull requests by improving quality gates and selection processes.
Impact: Better review processes reduce waste and improve the quality of shipped code, maximizing the value of AI-generated output.
-
Prioritize modernization of legacy codebases or implement advanced context engineering. Focus AI efforts on newer, AI-friendly codebases initially, while developing strategies to handle complex legacy systems.
Impact: Targeting high-impact areas first ensures quick wins and builds momentum, while long-term strategies address the challenges of legacy code.
Quotes
“Most people, they're working interactively with one to two agents. The most experienced engineers, they get stuck at four max.”
“We find ourselves now, with this massive data set, we've really cemented the understanding that there are real raw coding games, and we see them out in the wild.”
“We're at the point where engineering leaders also have to explain where are all the tokens going? Because it is a lot of money.”