AlgoFoxGet AlgoFox App

Courses

Agent Harnesses in Practice

A tool guide across six products: what a harness adds to a stateless model, how OpenClaw, Hermes, OpenCode, Oh My Pi, OpenRouter and Herdr each handle memory, skills, tools, routing and triggers, and how to choose one and lock it down.

Islands
6
Lessons
34
Exercises
194
DebugCode orderFill blankMultiple choiceTapTrace

Start the first lesson

  1. Island 1

    What a Harness Is

    A model is stateless. A harness adds five parts around it: gateway, session, memory, tools and triggers. Six products, each built around one part first, and how to choose one for a team.

    1. The Five Parts of a Harness5 exercises, 5 min
    2. Six Tools, One Shape5 exercises, 5 min
    3. Chatbot or Operator5 exercises, 6 min
    4. Checkpoint: What a Harness IsCheckpoint6 exercises, 6 min
    5. Boss: Choose a Harness for a TeamBoss8 exercises, 9 min
  2. Island 2

    Memory Designs

    The three ways a harness remembers: a small file loaded every session, a store searched from outside the window, and the flush that saves a fact before compaction rewrites the chat. What to keep, what to drop, and which scope a fact belongs to.

    1. The File in the Window5 exercises, 5 min
    2. Search Outside the Window5 exercises, 6 min
    3. Compaction and the Flush5 exercises, 6 min
    4. What to Store5 exercises, 5 min
    5. Checkpoint: Memory DesignsCheckpoint6 exercises, 6 min
    6. Boss: Remember a CustomerBoss8 exercises, 9 min
  3. Island 3

    Skills and Procedures

    Three kinds of remembered text and when each loads, the skills an agent writes for itself and the approval that gates them, and the skill hubs that turn a procedure into a dependency.

    1. Rule, Procedure, or Fact5 exercises, 5 min
    2. Skills the Agent Writes5 exercises, 6 min
    3. A Skill Is Supply Chain5 exercises, 6 min
    4. Checkpoint: Skills and ProceduresCheckpoint6 exercises, 6 min
    5. Boss: Grow a Skill LibraryBoss8 exercises, 9 min
  4. Island 4

    Tools and Safe Edits

    How a harness makes an edit safe to trust: anchors that prove the file is unchanged, a language server and a debugger that hand back evidence, a permission table where order decides, and subagents that return an object you can check.

    1. Edits That Prove Freshness5 exercises, 5 min
    2. The Editor in the Loop5 exercises, 5 min
    3. Permission Tables5 exercises, 6 min
    4. Typed Subagents5 exercises, 5 min
    5. Checkpoint: Tools and Safe EditsCheckpoint6 exercises, 6 min
    6. Boss: The Risky RefactorBoss8 exercises, 9 min
  5. Island 5

    Routing and Fleets

    One key in front of many models, the two routing layers, the provider rules that pin or spread, fallback lists and the auto router, and one person's view over a fleet of agents.

    1. One Key, Many Models5 exercises, 5 min
    2. The Provider Object5 exercises, 5 min
    3. Fallbacks and the Auto Router5 exercises, 6 min
    4. Many Agents, One Human5 exercises, 6 min
    5. Checkpoint: Routing and FleetsCheckpoint6 exercises, 6 min
    6. Boss: Survive a Provider Outage on Launch DayBoss8 exercises, 9 min
  6. Island 6

    Always-On Security

    Locking down an agent that never sleeps: the one Gateway that owns every channel and who may pair with it, heartbeats and scheduled jobs that run with nobody watching, the three ingredients on a chat channel, and the memory a stranger can write into every future session.

    1. The Gateway5 exercises, 5 min
    2. Triggers5 exercises, 6 min
    3. Three Ingredients on WhatsApp5 exercises, 6 min
    4. Memory Pollution5 exercises, 6 min
    5. Checkpoint: Always On, Locked DownCheckpoint6 exercises, 6 min
    6. Boss: Lock Down an Always-On AgentBoss8 exercises, 9 min