A coding agent can explore a repository, edit multiple files, run commands, and open a pull request while the developer handles something else. PR count rises immediately. Productivity only increases if the work passes review, reaches production, and justifies the human and token cost.
Agentic coding tools do more than suggest the next line. Claude Code can read a codebase, edit files, run commands, and create commits and pull requests. GitHub Copilot’s cloud agent can research a repository, develop a plan, make changes on a branch, and open a pull request. Cursor also supports local and cloud agents.
The output is a chain of actions: repository exploration, tool calls, edits, tests, retries, commits, and pull requests. That makes the agent run or assigned task a useful starting unit of analysis. Leaders still need to connect the run to the code that merged and the outcomes that followed.
A 2026 study of eight frontier models on SWE-bench Verified found that agentic coding tasks consumed about 1,000 times more tokens than code reasoning and code chat in the study setup. Runs on the same task varied by as much as 30 times, and higher token use didn’t consistently improve accuracy.
The findings aren’t a universal cost benchmark. They show why leaders need run-level usage, cost, and outcome data.
DORA’s current model measures change lead time, deployment frequency, failed deployment recovery time, change fail rate, and deployment rework rate. These metrics still show whether software delivery is becoming faster or less stable.
They don’t identify whether a change came from a human developer, inline AI assistance, or an autonomous agent. Deployment frequency can rise while agent-generated work creates longer review queues or more rework. The delivery metric is accurate, but the cause is unclear.
Segment agent-generated, human AI-assisted, and human-only work wherever the data supports it. This shows whether agent activity is improving outcomes or simply changing the source of the work.
Capture the agent, tool, model, owner, team, repository, task type, start and end time, and resulting pull request or associated commits.
Useful measures include agent runs started and completed, agent-attributed pull requests and commits, the share of merged work associated with agents, and runs without a named owner or business purpose.
Don’t combine agent activity with human AI usage in one total. They have different cost patterns, governance needs, and output characteristics.
A run that ends successfully may still produce unusable work. Measure the percentage of tasks that met acceptance criteria, passed required tests, reached review, and merged. Also capture retries, failed runs, human interventions, review time, and the amount of code substantially changed before merge.
These signals show whether the agent completed the task or handed the difficult part back to a developer. They also reveal where the workflow needs better context, narrower tasks, stronger tests, or clearer guardrails.
Track agent-generated changes for 30 and 90 days using code turnover, reverts, defects, security findings, incidents, and remediation work.
Compare agent output with matched human AI-assisted and human-only work in similar repositories and complexity bands. A higher merge rate means little if the code is repeatedly rewritten or creates production problems.
Larridin’s Developer AI Impact Framework connects AI attribution with throughput, quality, and cost. The code turnover measure shows how much committed code is substantially rewritten or deleted after merge.
Track PR cycle time, time to first review, review queue depth, change lead time, deployment frequency, change fail rate, deployment rework rate, and failed deployment recovery time.
An agent may shorten implementation time while extending review or testing. Another may generate fewer PRs but help complex work reach production faster. The goal is to see whether the bottleneck moved and whether total delivery performance improved.
Capture observed token usage, billed spend, tool calls, runtime, infrastructure, and any seat or platform costs associated with the agent. Add the human cost of review, debugging, remediation, and operational support.
Then divide the full cost by an outcome that matters, such as a merged PR that survives 30 or 90 days, a completed feature, or complexity-adjusted output that reaches production.
Larridin’s Token Spend & Insights attributes spend to agents, teams, workflows, projects, vendors, and models. Pair that cost view with delivery and quality data so a low-cost agent isn’t mistaken for a high-value one.
Start with a small number of defined workflows.
Monitor failed loops, cost spikes, and queue growth operationally. Use a monthly or quarterly leadership review to decide which agent workflows to expand, redesign, limit, or stop.
They show volume without context about whether the work merged, survived, improved delivery, or required extensive human repair. Pair them with durability, quality, review, delivery, and cost.
Compare the full cost of the agent workflow with the durable delivery value it creates. Include token and tool costs, human review, remediation, and operational support. Compare similar tasks rather than applying one benchmark across every workflow.
No. Measure the agent as part of a workflow and delivery system. A per-agent activity ranking can reward volume without value.
Give every agent a named owner and budget, separate agent spend from human AI usage, monitor consumption as it occurs, and set projected overage alerts. Larridin’s agent cost governance framework also connects agent activity to outcomes.
Larridin connects agent activity and spend with source control, delivery, quality, and business outcome data. Leaders can see which autonomous coding workflows create durable value, where human work is increasing, and which agents need tighter controls or a different role.
Book a discovery call to build your agentic coding productivity measurement model.