The Production Agent
How to ship agentic AI to production without getting paged at 3am.
Battle-tested playbooks from shipping multi-agent systems to production - orchestration, evals, security, MCP integrations, and the failure modes you don't read about in the launch posts.
The Production Agent
How to ship agentic AI to production without getting paged at 3am.
Sattyam Jain
Self-published, drafted in public
Chapters 1 and 2 are drafts you can read free on this site. Chapters 3 to 5 are outlined and not written yet. No chapter is on Substack yet, and nothing is for sale.
Start free. Chapters 1 and 2 are free to read here as drafts. No account needed.
Start with Chapter 1Table of contents
5 chapters · 2 free
- 01FREE22 min read
Designing the Production Agent
Orchestration patterns, agent contracts, and the deny-by-default starting line.
How to design an agent that survives contact with production. Tool inventories, capability contracts, deny-by-default surfaces, and the cost router pattern that takes simple problems off the frontier model.
Read draft - 02FREE18 min read
MCP Security & STDIO Posture
Why MCP STDIO is RCE-by-design and the sandbox patterns that fix it.
The Model Context Protocol unlocks composable agent tooling - and ships a critical-by-default vulnerability in its STDIO transport. CVE-2026-30623, what it actually means, and the three-layer sandbox pattern that keeps it from owning your host.
Read draft - 0331 min readOutlined
Agent Supply-Chain Audit
50-point audit for the upstream surface of an agent stack.
The upstream supply chain of an AI agent is model provenance + MCP server graph + eval-set lineage + secret rotation policy + audit-emit channels. 50 audit points, each with a concrete check and a remediation.
Not written yet - 0428 min readOutlined
Migrating to Claude Managed Agents
Step-by-step runbook for moving production workloads.
Claude Managed Agents shipped with Anthropic's nine-connector creative launch and ten finance-agent templates. Here's the migration runbook - control-plane wiring, capability-lease bindings, MCP connector hardening, and the rollback playbook for when a vendor surface shifts.
Not written yet - 0524 min readOutlined
Agent-on-Agent Commerce - The Honest Version
What works today, what's marketing, and what the next 12 months look like.
Anthropic Project Deal, Cloudflare Mesh, Cisco Agentic Workforce Identity, Okta for AI Agents - the agent-on-agent stack is real. Here's what actually ships in 2026 vs what's vendor positioning, and what an honest /api/quote agent contract looks like.
Not written yet
Why this book
Most LLM books teach you to call the API. Most agent posts celebrate a launch. Neither tells you what breaks at week 2 of production, what your audit-emit channel actually needs to capture, or why your MCP STDIO config just gave an attacker a shell on your box.
This is the missing manual: orchestration patterns, eval harnesses, capability-lease identity, MCP-security posture, supply-chain audit points, and the migration runbook for moving onto Claude Managed Agents. All of it distilled from shipping multi-agent systems to production, plus the open-source tooling (agent-audit-kit, agent-airlock, mnemo) that I use to keep them honest.
Chapters are drafted here, in public, before anywhere else. None are on Substack yet, and there is no paid tier or bundle to buy. Subscribe to The Production Agent (opens in a new tab) to hear when a chapter is ready.