The app for independent voices
I built my first Claude Managed Agent.
Surprised how easy it was.
You describe what you want in plain English. The platform generates the full agent config: model, system prompt, tools, MCP servers, permission policies. All in YAML you can edit.
I asked for an email reader that needs my approval before acting. It set permission_policy to always_ask, offered Gmail MCP, and suggested document skills — PDF, Excel, Word, PowerPoint.
Running in minutes.
Here’s what you need to know:
10 templates to start from: support agent, deep researcher, incident commander, data analyst, sprint retro facilitator. The incident commander comes pre-wired with Sentry, Linear, PagerDuty, and GitHub MCP servers — each with its own permission policy.
Each tool has its own permission level — always_allow, always_ask, or deny. You define autonomy boundaries per tool, not per agent. This is the guardrail layer.
MCP servers connect with a URL. Standard protocol, not custom integrations.
Environments are container configs: packages with version pinning, network rules, host whitelists. Sessions can mount GitHub repos and files into the container — that's how you get context in.
Agents are versioned. Every run attached to a specific version. Full debug timeline — every tool call, every error, timestamps. OTel export to your existing stack.
Currently cloud only — but the dropdown hints at local environments coming. That would be huge for enterprise.
Custom skills visible in the UI but not configurable yet. Upload via Skills API. Still beta.
—
What's coming (research preview, request access):
Outcomes: define what "done" looks like as a rubric. A separate grader evaluates in its own context window. Agent iterates up to 20x until satisfied.
Multi-agent: declare callable_agents in YAML. Engineering Lead delegates to Reviewer + Test Writer. Each versioned independently.
Memory stores: persistent memory across sessions. Per-user, per-team, or per-project. Up to 8 stores per session. Full audit trail with versioned rollback.
—
It doesn’t replace Claude Code. No hooks. Local gives you direct filesystem, git, and your dev tools. Managed runs in isolated containers — but you can mount GitHub repos and files into each session.
Local for development. Managed for production tasks you want to run headless.
Price: $0.08 per session-hour of active runtime. Idle time doesn't count. Model tokens on top at standard API rates.
—
This is their AWS moment.
—
You can download all my infographics here: go.productcompass.pm/le…
Want to master building with Claude? On May 9 we're launching Claudathon for PMs. Four weeks, one real AI product. No coding required. Last spots: go.productcompass.pm/cl…
