SaaS Platform Team Ships Features 3x Faster With 12 Production Claude Agents

An 8-person engineering team at a B2B SaaS company wanted to move faster without hiring. They’d seen demos of Claude Code doing code review, writing tests, and updating documentation. The pitch was obvious: use AI agents to multiply the output of their small team.

The Problem

They started with two agents running directly against the GitHub API and their AWS environment. Both used the same developer’s personal credentials — his AWS access key, his GitHub token with admin access.

Within three weeks, they had their first incident. One of the agents, tasked with “clean up the staging environment,” interpreted that broadly and deleted a CloudWatch dashboard they’d spent a week building. Nothing critical — staging is staging — but it planted a question: what would this agent do with the same permissions in production?

They also hit a practical limit at two agents. Adding a third meant manually managing another set of credentials, another set of instructions. There was no unified view of what any of the agents were doing. When one agent’s code review conflicted with another agent’s refactoring, nobody knew until they hit a merge conflict.

They needed infrastructure before they could scale.

The Solution

The team onboarded Sentrely and restructured their agent fleet properly.

Each of their 12 agents has a specific job and a scoped policy:

• code-review-agent: read-only GitHub access, can post PR comments, cannot push
• test-generator-agent: read repo, write to feature branches only
• deploy-agent: specific ECS service deployment, Slack notification on completion, Slack approval required for anything touching production
• docs-agent: read codebase, write to /docs directory only

The A2A messaging setup solved their coordination problem. When code-review-agent posts a review and requests changes, test-generator-agent picks up the task automatically without human orchestration.

The unified dashboard shows all 12 sessions in real time. When an agent is stuck — trying to do something outside its policy — it shows up immediately rather than silently failing.

The Results

3x feature velocity. Not because the agents are smarter than before — they’re the same Claude. Because the team trusts them enough to run 12 of them on real work, and the coordination infrastructure makes them work together rather than against each other.

Zero unreviewed merges to main in 6 months. The deploy-agent policy requires a human click on every push to main. It’s never been bypassed. The team considers this a meaningful safety record.

2 weeks to first production agent. From “we should use Claude agents” to a scoped, audited agent running on real code.

The CTO’s take: “I used to worry that AI agents would be a liability. Now I worry about the teams that aren’t using them with proper controls.”