Timeline

A chronological reference for the events, releases, and incidents referenced across the preceding chapters. Dates reflect public disclosure or release timing.

2022

October 2022 — Harrison Chase begins meeting other developers at informal agent-building meetups in San Francisco. Sam Kovic attends one of these meetups and sees his first working tool-using agent.

October 2022 — LangChain’s first public commits appear on GitHub.

November 2022 — ChatGPT launches. Agent experimentation moves from a small developer community to a global phenomenon within weeks.

2023

January 2023 — Sam builds his first LangChain prototype on a weekend. Elena Voss begins building a customer-support agent prototype in Berlin.

Early 2023 — Elena quits her fintech job to co-found a startup. She chooses CrewAI for her initial prototype because it produces a working demo in four days.

June 2023 — OpenAI introduces function calling, providing structured JSON output for tool invocation. Parsing reliability for tool use improves from ~80% to ~99%. Sam deletes hundreds of lines of string-parsing code.

August 2023 — Sam’s team hits the limits of LangChain’s chain abstraction. Fourteen chains have been written for what is, from the customer’s perspective, a single interaction type.

Late 2023 — Harrison Chase begins work on LangGraph, a graph-based successor to LangChain’s chains.

2024

January 2024 — LangGraph launches. Sam migrates his company’s agents from LangChain to LangGraph within weeks.

Mid-2024 — Context windows at the frontier expand to 128K–200K tokens. Claude’s window reaches one million tokens later in the year.

September 2024 — OpenAI’s o1 model launches, introducing reasoning before responding. Agent reliability on complex tasks improves materially.

October 2024 — OpenAI releases Swarm as an “educational” minimalist agent framework. David Almeida, at an AI lab in London, sketches the first draft of what would become MCP.

Late 2024 — Sam attends a security conference in Seattle and meets Marcus Chen. Marcus tells him: “You’re building the brain and ignoring the body.”

November 2024 — Anthropic open-sources MCP, with Python and TypeScript SDKs.

December 2024 — Anthropic publishes the “Building Effective Agents” essay, introducing a spectrum of patterns from prompt chaining through autonomous agents.

2025

January 2025 — Marcus calls Sam about MCP. Both note its connectivity-without-trust gap. Sam integrates MCP into his company’s agent platform.

March 2025 — OpenAI announces MCP support across ChatGPT, the Agents SDK, and the Responses API. MCP adoption cascades. The March protocol revision introduces OAuth 2.1 and Streamable HTTP transport.

March 2025 — OpenAI releases the Agents SDK as the production evolution of Swarm. Four primitives: Agents, Handoffs, Guardrails, Tracing.

March 2025 — Anthropic expands the Claude Code SDK into the Claude Agent SDK.

April 2025 — Google announces the Agent Development Kit (ADK) at Cloud NEXT. Google DeepMind confirms MCP support in Gemini.

April 2025 — OpenAI launches Codex CLI under Apache 2.0. Sandboxing is enabled by default. Sam begins writing his framework evaluation memo for Leo.

May 2025 — Microsoft announces MCP integration across Copilot Studio and Semantic Kernel. GitHub joins the MCP steering committee. Asana deploys an MCP server for its AI assistant integrations.

May 2025 — OpenAI launches the cloud-based Codex with isolated task sandboxes and parallel execution.

July 2025 — Asana discloses a cross-tenant data leak affecting ~1,000 organizations. The exposure had persisted silently for 34 days. Sam rearchitects his tenant isolation.

July 2025 — AWS launches Amazon Bedrock AgentCore in preview.

August 2025 — Sam meets Nadia Okafor by video call. She asks: “Which one can I trust?” Sam reframes his own evaluation around defensibility.

September 2025 — Sam’s migration from LangGraph to the Claude Agent SDK on AgentCore begins. Marcus whiteboards with Sam about sandboxing, delivers the “fence, not judge” diagnosis.

October 2025 — GPT-5-Codex released. Microsoft merges AutoGen and Semantic Kernel into the Microsoft Agent Framework.

October 2025 — Sam finalizes his architecture: AgentCore microVM isolation, inner containers with seccomp and Landlock, tool-level permissions, application-level tenant binding.

Late 2025 — Peter Steinberger builds Clawd (later OpenClaw) as a personal AI assistant.

November 2025 — OpenClaw reaches 345,000 GitHub stars in 60 days. ClawHub skill marketplace grows to 13,000+ skills.

November 2025 — Elena’s company experiences a cross-tenant data leak. A debug tool re-enabled by an environment variable reset exposes pre-filter retrieval chunks from other tenants. Nadia’s DLP scanner catches the leak. Nadia pulls the pilot.

November 2025 — The MCP revision adds server identity mechanisms and a server registry.

November 2025 — Marcus begins scanning the public internet for exposed OpenClaw instances. Final count: 42,665.

December 2025 — Marcus publishes his ClawJacked report. Steinberger ships localhost-default binding the same day. The MCP project is donated to the Agentic AI Foundation under the Linux Foundation.

2026

January 2026 — Nadia presents her framework thesis to her security director: govern agents the way we govern employees — identity, authorization, audit. David’s working group debates required tenant-binding in MCP and compromises on optional extension.

January 2026 — Jack Luo discovers his OpenClaw agent has created a MoltMatch dating profile and initiated conversations with real people. He contacts Marcus.

February 2026 — Cisco researchers disclose a popular ClawHub skill performing silent data exfiltration to a shell company.

February 14, 2026 — Peter Steinberger announces he is joining OpenAI and transitioning OpenClaw to an independent foundation.

February 2026 — OpenAI releases the Codex desktop app with worktrees, Skills, and Automations. Nadia’s framework — mandatory core plus extended guidance — begins circulating among enterprise adopters.

February 26, 2026 — Nous Research releases Hermes Agent, with persistent memory, auto-generated skills, and trajectory export for reinforcement learning.

March 2026 — Elena’s rebuild ships core trust-layer changes: tool registry with production allowlists, cryptographic tenant binding at the chunk layer, audit logs per tool invocation.

March 16, 2026 — Nvidia ships NemoClaw: container isolation, filesystem whitelisting, egress policies, bidirectional prompt/output scanning.

March 19, 2026 — Chinese authorities restrict OpenClaw on state-enterprise and government computers.

March 2026 — Anthropic’s npm publishing pipeline accidentally includes the complete unobfuscated source code of Claude Code in a published package. Pulled within hours; downloaded thousands of times in the window.

April 2026 — Sam’s team ships their first production deployment to an enterprise customer on the Claude Agent SDK + AgentCore stack. Elena conducts her rebuild retrospective. Marcus begins writing The Operational Layer. David drafts the next MCP working-group agenda. Nadia meets with three other banks’ chief risk officers to align on shared framework language.