When a software company saw agent cycle time spike 15x, the dashboard said failure

• Systems integratedGitHubJira
• Agents instrumentedClaude CodeCodexCopilotCursor
• TeamFewer than 1,000 developers
• StakeholderCTO
• ProcessAI software development lifecycle

Highlights

• Adoption tracked by tool and by model: no agent sessions in May, rising to 67% of the platform team by July
• Every agent session tied to the work it touched, down to AI code share on each pull request, ranging from 8% to 96%
• Human-only cycle time on the platform team fell 49% from May to August, while agent-assisted work grew from 11% to 54% of completed work
• On support work, agent-assisted cycle time fell 68% from its June peak and converged with human-only work

Situation

This private equity-backed software company is growing both organically and through acquisition. It had started putting agents into its engineering workflows, with Claude Code, Codex, Copilot, and Cursor in use across two workstreams: platform engineering, and support and maintenance. Adoption was early and uneven, and the CTO wanted it to grow.

Challenge

Growing adoption meant knowing where it stood: which developers used agents, with which tools, and on what work. It also meant knowing what agent-assisted work looked like once it reached GitHub, how much of the merged code agents wrote, and whether that work finished faster than the same team's work without an agent.

The agent tools showed usage, not who drove it or what it changed. A metric moving the wrong way with no way to explain it is how AI programs get cancelled.

Solution

Bloomfilter connected to the Jira and GitHub instances the company already ran and mined the event streams they already produced. No new workflows, no manual mapping. End-to-end flow became visible, including where work moved backward.

Bloomfilter then overlaid agent activity from all four tools on that same structure. A developer binds a session to a Jira ticket, and the session inherits the pull request, build, and deployment that ticket already correlates to. Capture runs through an open-source plugin, with prompt contents excluded and nothing leaving the boundary the company had already approved.

With sessions tied to work, three comparisons became possible: adoption by tool and model, AI code share on every pull request, and cycle time for human-only work measured against human-plus-agent work.

Adoption became a number the CTO tracks

No agent sessions in May. Adoption climbed through June and reached 67% of the platform team by July. Average turns per session more than doubled between June and August and held there. Engineers were working with the agents, not firing one-shot prompts. File-edit volume kept climbing through July, the month the team stopped experimenting. In August, adoption dipped to 50% and sessions fell, yet agent-assisted work rose to 54% of everything the team completed.

The metric that looked like a problem

In August, agent-assisted cycle time jumped from 1.3 days to 21.0 days. A 15x increase in one month, on the flagship AI initiative. On a usage dashboard, that is where the program ends.

With sessions tied to work, the company could see the other half of the picture. In the same month, human-only cycle time dropped to its lowest point in four months, and agent-assisted work went from 22% to 54% of everything the team completed. The agents had not slowed down. They had been given more complex work that takes more time.

By August, agents were working on core platform architecture

The team did not run more sessions. It put agents on harder problems. The story point range on agent-assisted work climbed month over month, and the nature of the work changed with it.

By August, agents were co-authoring security infrastructure and multi-week architectural features. The spike measured harder work, not slower agents.

When an agent writes 96% of a pull request, the engineer directs the work

AI code share ranged from 8% to 96% depending on the task. The highest shares landed on scoped fixes and stabilization work, where the agent could carry the change end to end. On new functionality and test coverage, engineers wrote more of the code themselves. On the high-share pull requests the agent was the primary author and the engineer was the reviewer and the decision-maker. Bloomfilter measures that split per pull request, so the ratio is a fact about the merge rather than a claim about the tool.

On support work, the same instrumentation showed a different shape

The support and maintenance workstream followed a different curve. Agent-assisted cycle time fell 68% from its June peak. Task volume rose 53% over May. Velocity rose 22%.

By August, agent-assisted cycle time (11.1 days) had converged with human-only cycle time (10.0 days). The agent was carrying work of comparable complexity to what the team handled without it. Convergence is what maturity looks like, and it is only visible when both numbers come from the same delivery record.

What changed

Seeing the spike meant the company could hold course through it. Seeing what the agents were working on meant the CTO could steer adoption toward the hardest problems instead of mandating it across the board. The lesson was not that agents make everything faster. It was that agents let the strongest engineers take on the problems they had been routing around.

About the data. Unattributed agent sessions are excluded, so actual AI usage was likely higher. Adoption rate is measured against the platform team, not the full engineering organization.