80% of ops.ai’s agent cost was context passed between agents. The sessions showed it. The invoice never could.
ops.ai is an AI-native software consultancy that designs, builds and runs enterprise-scale products for its clients, from high-traffic e-commerce platforms to mobile apps, as well as its own feature-flag platform, Toggly. Nearly all of that work is shipped by autonomous Cursor and Codex agents. Promptster showed ops.ai where those agents’ tokens actually went, and what its way of working actually ships.
pull requests completed in September by the founder’s count
production deploys a day, with no production outage in two months
unique visitors a month across the products ops.ai runs
The story
The problem
ops.ai runs agents for up to 16 hours without a single prompt. They follow a “constitution” of delegation rules and about 200 pages of design specs. The team could see the runs were long, but not what the agents spent their time and tokens on.
What Promptster found
Promptster’s delegation breakdown showed agents opening large numbers of sub-agents and handing their full context to each one. Contradictions in the constitution were driving it. That handoff was about 80% of the cost, and it confused the agent reviewing the work.
What changed
ops.ai changed how its agents split up work, size it and iterate on it, and removed the contradictions from its rules. By the founder’s account, spend dropped significantly.
Before and after the change
- 6×More work finished. ops.ai closed 611 Linear issues in September, up from 104 in July, before the change.
- 5×Faster, at six times the volume. The median issue went from started to done in 4.8 hours, down from 23.5.
“A lot of the process improvements I got by using Promptster. I wasn’t aware of how much waste was going into context getting passed around, and the agents splitting at the wrong place.”
The next workflow change
What Promptster found
Promptster classifies every session by how the work was driven: from a written spec, from a plan, or by ad-hoc prompting. It then follows each one through to a merged PR.
The result
At ops.ai, coding sessions that started from a written spec ended in a merged PR 75% of the time. Ad-hoc sessions shipped 34% of the time. The gap held when comparing sessions of the same size.
What ops.ai does next
Start every coding task from a spec, and watch the ship rate on the same dashboard that caught the delegation problem.
The ROI questions ops.ai can now answer
- ~$10What does each shipped change cost? About $10 of AI usage per merged PR, and about $12 per Linear issue closed, across 670+ PRs and 551 issues Promptster matched in 30 days.
- 80%Is the AI’s output any good? 80% of AI-written lines were still in the code 30 days later. 95% of CI runs passed, and no merged PR was reverted.
- 59%Where is the money going? One model was 59% of the month’s spend, and a single 14-day agent session cost $2,506. That points ops.ai at its next saving: a cheaper model for routine execution.
Last 30 days to Sep 28, 2026. Cost prices every token at the vendor’s API list rate, whatever the plan. Promptster’s PR and Cursor usage counts are floors.
What Promptster sees at ops.ai
- Every rail, passively. Cursor and Codex sessions, including the scheduled agents, with no workflow change.
- Where spend goes. Token usage, sub-agent calls, skills and MCP servers, and top spenders by session, next to how the work was driven.
- Where agents wait. Hours agents sat waiting on a human, the next bottleneck after cost.
- Now live, built with ops.ai: an MCP server ops.ai’s agents query for the same data as the dashboard, to tune their own spend.
Why it matters
“Most companies are using more than one model and more than one harness, so if one goes down they can switch. Tracking spend across all of those is a big thing, and so is figuring out which employees are using it effectively.”
Privacy by design
Open-source capture
The capture tool that runs on each engineer’s machine is MIT-licensed and public at github.com/pa-arth/promptster-teams-cli. Anyone at ops.ai can read exactly what it collects.
Code never leaves the machine
Before anything is sent, the tool strips diffs, file contents, command output and the model’s reply text, and redacts secrets. Our database also rejects any row that carries source code.
Check it yourself
Every event is written to a local log before it is sent, already redacted and signed. That log is exactly what leaves the machine, and anyone can open it and read it.
ops.ai kept its own Cursor and Codex subscriptions, with nothing routed through a Promptster proxy. Session data is kept for no more than 90 days, deleted on request, and never sold or used to train models. Every figure here is an aggregate ops.ai reviewed. How we handle data.
One squad or the whole org. Runs on the Claude Code, Codex and Cursor seats your engineers already have.
Promptster for Teams →See where your agents’ tokens go. Free cost audit, 5 minutes.
Run the cost audit