Search the Practicum

Type two letters · ↑↓ select · Enter open · Esc close

Type at least two letters.

Autonomous agents: runtimes and frameworks

This is a live Practicum page. Verified: 2026-08-11. This tool category changes every month, so check the date. This page is for readers who are ready to work through the technical details. Most readers can start with built-in assistants such as Gemini Spark or ChatGPT Work with its cloud browser. See simple agents. This page covers a separate category: autonomous agents as runtimes, their architecture, and their connection to the harness in Chapter 6.

A chat answers a question, then loses the thread. An autonomous agent is a runtime, not a chat. A runtime is a program that keeps running. It plans steps, calls tools, remembers context across sessions, and can improve over time. Here, the English word "agent" means that whole working setup. It is not a smarter answer box. It is software that acts.

The architecture: what these agents are made of

The names vary, but the basic structure has five parts:

  1. A model as the brain. This can be any LLM, or large language model, that works with language. Good runtimes are model agnostic: you can connect Claude, GPT, or a local model and replace the brain without rebuilding everything.
  2. A plan, act, observe, check loop. The agent breaks a goal into steps, takes one step, reads the result, and adjusts. This loop is what separates an agent from a single answer.
  3. Tool calling. The model goes beyond describing an action and calls tools such as files, a calendar, a browser, code, or messaging apps. A standard tool interface sits at the center of most runtimes.
  4. Persistent memory. Context remains available across sessions. Many runtimes store it in ordinary Markdown or YAML files on your computer. You can open and inspect that memory instead of trusting a black box.
  5. A skill loop that compounds. After solving a hard task, the agent records a reusable skill: the procedure, common mistakes, and checks. It does not have to learn the same lesson again. This is what our skills/ folder does: a correction becomes a skill that the system can reuse.

How this connects to the book's harness (Chapter 6)

In Chapter 6, "The Human as a System," a harness is the working setup around the model and around you: context, tools, checks, and boundaries. An autonomous agent uses the same harness, but moves part of it into software. You still own and review the outcome. The runtime takes over planning and routine steps.

The industry has started using similar language. The optional OpenClaw.NET gateway describes an agent's plans as Harness Contracts, which are inspectable work plans. Its safer mode uses Plan-Execute-Verify with Evidence Bundles, or receipts for each step. This is the machine version of the book's "show the receipts" rule. Do not ask people to trust the agent. Show what supports its conclusion and what another person can check.

Real examples, with sources

This is an open-source, MIT-licensed agent that you can host yourself. You can keep the runtime and working files on your own system, but a cloud model provider will still receive the requests you send to its model. Hermes keeps long-term memory about your projects and writes its own skills after difficult tasks. Those skills use a format compatible with the open agentskills.io standard, the same idea used in our skills/ folder. One gateway can reach you through Telegram, Discord, Slack, WhatsApp, Signal, or a command-line interface. The agent is model agnostic and supports Nous Portal, more than 200 models through OpenRouter, OpenAI-compatible endpoints, and local vLLM. Installation uses one command. -> hermes-agent.org · GitHub

OpenClaw is an open-source autonomous agent, previously called Clawdbot. Its core combines an agent loop, persistent memory, tool calling, and chat integrations. It stores memory and skills as plain Markdown and YAML files in its working folder (~/.openclaw), where you can inspect them. Its .NET gateway adds Harness Contracts, Evidence Bundles, and Plan-Execute-Verify mode for risky actions. OpenClaw takes more work to set up than Hermes. -> docs.openclaw.ai · GitHub · OpenClaw.NET

n8n and Zapier connect apps in fixed workflows without a planning brain. They are inexpensive and dependable for repeated tasks. See simple agents. ChatGPT Work with its cloud browser and Gemini Spark give you a managed agent without self-hosting. They are easier to start with, but your data goes to the provider.

Who may need one, and why

  • A specialist with a repeated cycle, such as sorting email, preparing a morning brief, or monitoring sources, can assign one self-hosted agent to that cycle. Its skill loop can improve the process over time.
  • An analyst or product manager can schedule an agent to collect data from several sources and prepare a draft with receipts. The person checks the work and owns the conclusion.
  • Someone who wants more control over data can use a self-hosted option such as Hermes or OpenClaw. Working files can stay on their system, but a cloud model provider still receives requests sent to its model.
  • You may not need this at all. A built-in assistant is enough for a one-off task or a first experiment. Do not try to build an "agent company" yet. That belongs in Volume 2. Start with one cycle.
Discipline

An agent needs boundaries and checks. Do not give it access to sensitive systems such as email, payments, or personal data until you understand what leaves your device and where it goes. Ask for receipts that show what supports each conclusion. One tested cycle is more useful than a plan to automate your entire life. Do not trust. Verify.

For the personal, simpler view, see simple agents. For the tool overview, see the tool map. Version: 2026-08-11.

Sources (verified 2026-08-11)

Primary sources for autonomous agent runtimes and frameworks:

Runtime features change quickly. Check the current product before relying on a feature, and treat vendor demos as claims.

Autonomous agents: runtimes and frameworks