Agents, plainly · Part 2 of 3 · Aug 2026 · Part 1

From chat to a staff of agents

The fastest way to feel the shift: in 2024 the best AI product was a glorified search box. In August 2026 the loudest launch is a team of agents with their own computer that keeps working at 2am. That's four eras in two years.

The chat era — 2022 to 2024

ChatGPT, Claude, Copilot: you ask, it answers. The interface was a search box and the model was the product. Everyone sat equally behind the wheel, which is why the gap I keep writing about didn't exist yet.

The coding-agent era — 2025

Claude Code began as one engineer's side project in his first month at Anthropic and defined a category: give an agent the files, a terminal, and a goal, and it builds. MCP gave models a standard plug for tools. The era happened in public, inside developer tools.

The personal-agent era — 2026

OpenClaw (≈390k GitHub stars) and Hermes Agent (≈237k stars, MIT) brought the agent out of the terminal: one agent, one memory, across Telegram, Discord, Slack, WhatsApp, Signal, and email. The chat products grew up too — Claude Cowork runs Claude Code's engine aimed at knowledge work, and ChatGPT Work pulls in your team's tools. And the harness layer got serious: DeepSeek Harness shipped its developer preview — everything is a plugin — while Pi shipped as a minimal terminal harness.

The orchestration era — now

Grok Bot (SpaceXAI, formerly xAI) launched early beta on 11 August 2026: your team of always-on agents, each with its own computer, working inside apps 24/7, priced for businesses. The promise is no longer “one agent answers.” It's “a staff that keeps working while you sleep.”

How they stack

How the layers stack. Each layer composes the one below — most people will only ever touch the top two, which is exactly the point of the bottom two.

Raw models are the foundation — GPT-5.x, Claude, Grok, Gemini, DeepSeek V4 in Flash and Pro. On top sits the harness: the loop, memory, tools, and permissions that turn a brain into a worker. That's where DeepSeek Harness and Pi live, and where I spend my days. Above that are the agents people actually run — OpenClaw, Cowork, ChatGPT Work, Claude Code. And at the top, the newest layer: orchestration of many agents at once, the always-on team.

Two notes worth keeping. First, the same loop runs at every level — memory, tools, a loop, a guardrail — the packaging is what changes. Second, the models at the bottom are becoming commodities (DeepSeek just shipped V4 Pro at scale for a fraction of the cost of its peers). The leverage has moved up the stack, into the harness and the orchestration. That's not an accident; it's where the builders went.

Own the harness and you ride every product wave that comes — or you build one of your own. Part three: what professionals actually do with all of this now.

Sources: xAI — Introducing Grok Bot · VentureBeat on Grok Bot · Claude Code history · OpenClaw · Hermes Agent · Pi · DeepSeek Harness · Reuters on V4 Pro