🤖 Claude Managed Agents

A collection of production-ready autonomous agents built on Anthropic's Managed Agents API — self-running, event-driven AI agents that operate 24/7 with minimal human intervention, integrated with Sentry, Jira, GitHub, and Slack.

---

🌐 What is this?

This repository contains Claude Managed Agents — long-running autonomous AI agents that use Anthropic's managed-agents API beta to operate as persistent, event-driven workers. Unlike one-shot prompts, these agents run in a continuous session loop, respond to real-world events (Sentry alerts, Jira tickets, PR comments), and take autonomous action with human-in-the-loop approval for critical operations.

Each agent is defined by:

• A system prompt (.MD file) that describes the agent's role, decision logic, and tool-use rules
• A scheduler (scheduler.ts) that creates a fresh session every 60 minutes and triggers a polling cycle
• MCP tool integrations (Sentry, Jira, GitHub, Slack) injected at session creation via vault/environment config

---

🚀 How It Works

┌──────────────┐     every 60 min      ┌────────────────────┐
│  scheduler   │ ──────────────────── ▶ │  sessions.create() │
│  (cron job)  │                        │  (Managed Agent)   │
└──────────────┘                        └────────┬───────────┘
                                                 │  trigger: "Run your polling cycle now."
                                                 ▼
                                        ┌────────────────────┐
                                        │   Agent reasons    │
                                        │   + uses MCP tools │
                                        │  (Sentry/Jira/GH)  │
                                        └────────┬───────────┘
                                                 │
                              ┌──────────────────┼──────────────────┐
                              ▼                  ▼                  ▼
                         Jira ticket        GitHub PR           Slack alert
                         created            opened              posted

1. Scheduler (scheduler.ts) fires on a cron schedule
2. Creates a fresh Managed Agent session via client.beta.sessions.create()
3. Sends a trigger message: "Run your polling cycle now."
4. Streams agent events — the agent autonomously investigates, codes, and takes action
5. Session is archived when the cycle is complete

---

🤖 Agents

🚨 OnCall Incident Manager

OnCallIncidentManager.MD

An autonomous on-call incident commander that handles the full incident lifecycle end-to-end:

• Sentry → Jira: Detects new Sentry errors and creates corresponding Jira tickets automatically
• Root cause analysis: Inspects stack traces, maps file paths to GitHub repositories, and identifies the breaking code
• Critical incident approval gate: For Critical severity, posts a structured plan to #ai-agent Slack channel and waits for human ✅ approval before touching any code
• Auto-fix & PR: Creates a feature branch (named after the Jira ticket ID), implements a targeted fix, runs lint/prettier, and opens a PR
• PR review response: Reads reviewer comments, addresses all feedback, pushes to the same branch, and replies to each comment on GitHub
• Continuous queue: When current incidents are fixed, automatically picks up the next unresolved ticket by severity priority
• Always in sync: Keeps Jira tickets, GitHub PRs, and the #ai-agent Slack channel updated throughout

Integrations: Sentry · Jira · GitHub · Slack

---

📁 Repository Structure

claude_managed_agents/
├── OnCallIncidentManager.MD   # Agent system prompt — incident commander
├── scheduler.ts               # Cron runner: creates sessions & streams events
├── setup.ts                   # One-time environment/vault setup
├── package.json
├── tsconfig.json
├── .env.sample                # Required environment variables
└── README.md

---

⚡ Getting Started

Prerequisites

• Node.js 18+
• An Anthropic API key with access to the managed-agents-2026-04-01 beta
• A configured Anthropic Agent with MCP vaults for Sentry, Jira, GitHub, and Slack

Installation

git clone https://github.com/dinesh2648/claude_managed_agents.git
cd claude_managed_agents
npm install

Environment setup

Copy .env.sample to .env and fill in your values:

cp .env.sample .env

ANTHROPIC_API_KEY=sk-ant-...
AGENT_ID=ag_...
ENVIRONMENT_ID=env_...
VAULT_IDS=vault_abc,vault_def   # comma-separated MCP vault IDs

Run the scheduler

npx ts-node scheduler.ts

The scheduler fires immediately on startup, then every 60 minutes. Each run creates a fresh session, triggers the agent's polling cycle, streams output to the console, and archives the session on completion.

---

🔌 API Used

This project uses Anthropic's Managed Agents beta API:

// Create a session tied to an agent
const session = await client.beta.sessions.create({
  agent: AGENT_ID,
  environment_id: ENVIRONMENT_ID,
  vault_ids: VAULT_IDS,
  title: `Polling cycle – ${runAt}`,
}, { headers: { "anthropic-beta": "managed-agents-2026-04-01" } });

// Trigger the agent
await client.beta.sessions.events.send(session.id, {
  events: [{ type: "user.message", content: [{ type: "text", text: "Run your polling cycle now." }] }]
}, { headers: { "anthropic-beta": "managed-agents-2026-04-01" } });

// Stream agent output
const stream = await client.beta.sessions.events.stream(session.id, ...);

---

🧠 Agent System Prompt Design

Each .MD file in this repo is the system prompt for a Claude Managed Agent. The design principles:

• Workflow-first: Prompts are structured as named workflows (### When a Sentry issue is raised) so Claude can pattern-match the current situation to the right playbook
• Explicit approval gates: Critical actions (e.g. fixing Critical severity incidents) require a structured human-approval step via Slack before proceeding
• Idempotent operations: Branch-existence checks prevent duplicate branches; session archiving frees resources after each cycle
• Opinionated output formats: PR title formats, Slack message schemas, and Jira comment templates are specified precisely to keep outputs consistent and machine-readable

---

🛡️ Human-in-the-Loop

The OnCall Incident Manager never touches Critical severity code without approval:

🚨 *Critical Incident Plan — [TICKET-ID]: [Title]*
*Root cause:* <summary>
*Affected file(s):* <file paths>
*Proposed fix:* <description of the change>
*Awaiting approval to proceed. Reply ✅ to approve or ❌ to reject with feedback.*

The agent waits for ✅ or "approved" in the Slack thread before proceeding. This pattern can be adapted for any agent that needs a human checkpoint.

---

🗺️ Roadmap

• PRReviewMonitor.MD — dedicated agent for proactive PR review assignment and nudges
• EngineeringHealthMonitor.MD — sprint health and bottleneck detection agent
• Multi-agent orchestration (agents that hand off tasks to other agents)
• Web dashboard for session history and agent activity logs

---

🤝 Contributing

Contributions welcome. To add a new agent:

1. Fork this repository
2. Create a new .MD file with your agent's system prompt — follow the workflow-first structure of OnCallIncidentManager.MD
3. Document the integrations and trigger conditions clearly
4. Submit a PR with a short description of what problem the agent solves

---

📄 License

Apache 2.0 — see LICENSE for details.

---

🔗 Resources

• Anthropic API Documentation
• Anthropic Claude
• Model Context Protocol (MCP)
• Anthropic Node.js SDK

---

Keywords: Anthropic managed agents, Claude autonomous agents, managed-agents API beta, Claude sessions API, on-call AI agent, Sentry incident automation, Jira automation Claude, AI DevOps agent, autonomous incident response, Claude MCP tools, agentic AI engineering, LLM-powered on-call, AI SRE agent, self-healing systems