Skip to content
Digital Nestby Ulises Jeremias

Persistent workspace layer · powered by Agent Toolkit

Your coding agent sessions are temporary.
Your workspace should not be.

Agentic Harness is a workspace scaffold, not another runtime. It holds the knowledge, personas, packs, project collections, and run history that Agent Toolkit's engines read and write — so context survives across sessions, repositories, and clients.

GitHub repository ↗How the runtime works
Persistent workspace state and a temporary coding sessionKnowledge, personas and project context feed the Agent Toolkit runtime. A separate illustrative coding session can end while those workspace cores and their runtime connection remain.Kknowledge/Cpersonas · packsPprojects/DEMO SESSIONactive · temporarySESSION ENDEDworkspace remainsAGENT TOOLKITloops · jobs · memory
An illustrative model, not a live agent session. The session is temporary; knowledge, personas and project context remain connected through the runtime.

Mental model

The session disappears. The workspace does not.

Agent Toolkit owns the engines — loops, queues, swarms, memory commands. A harness workspace owns the concrete state those engines read and write: knowledge, personas, packs, project collections, and run history.

What persists here

  • Knowledge

    Processes, learnings, todos, and patterns — version-controlled and searchable.

    • agent-toolkit memory search "topic"
    • agent-toolkit memory inject
    • agent-toolkit memory add --type learning "..."

  • Context

    Personas scope how the AI works; packs carry per-client and per-project context.

    • personas/implementer.md · reviewer · researcher · architect · writer · debugger · tester
    • agent-toolkit workspace use-persona reviewer
    • agent-toolkit workspace load packs/my-client.yaml

  • Projects

    Indexed repositories behind projects/ symlinks, cloned on demand into repos/.

    • agent-toolkit project clone owner/repo
    • agent-toolkit project list
    • .workspace.yaml org + clone policy

Where Toolkit execution meets this state

  • Loops — Definitions live in the Toolkit; each workspace owns its loop instances under loops/<name>/ with STATE.md checkpoints and report.md artifacts.
  • Jobs — agent-toolkit devcompanion provides queue semantics; this workspace holds templates/jobs plus the actual job state and plan artifacts.
  • Sessions — Sessions are temporary on purpose. agent-toolkit workspace context rehydrates persona constraints, loaded packs, and knowledge at the start of every one.

Quick start

Clone it. Initialize. Put it to work.

Every command runs through the Toolkit CLI you already installed. The workspace only owns the state those commands act on.

  1. Scaffold a workspace
                  git clone https://github.com/ulises-jeremias/agentic-harness ~/.agentic-harness
  2. Initialize
                  cd ~/.agentic-harness && ./scripts/workspace-init.sh
  3. Index a repository
                  agent-toolkit project clone owner/my-repo
  4. Start a read-only loop
                  agent-toolkit loop init daily-triage && agent-toolkit loop run daily-triage
  5. Review what it found
                  agent-toolkit loop status && agent-toolkit loop audit daily-triage

Alternative scaffold: agent-toolkit workspace init --dir my-workspace creates the same structure anywhere.

Terminology

Three things called “pack”

Same noun, three precise meanings — defined in the Toolkit's concept docs.

  • Solution packsToolkit repo · packs/

    Docs-only workflow templates (oss-maintenance, engineering-workflow…). Not loaded by the compiler.

  • Workspace packsThis workspace · packs/*.yaml

    Concrete client/project context loaded with agent-toolkit workspace load.

  • Loop packsThis workspace · packs/*.yaml

    Runtime overrides merged by agent-toolkit loop run --pack.

Boundaries

Who owns what

The normalization was about boundaries. Each concept has exactly one owner; everything else installs, uses, or simply stays out of the way.

ConceptWorkstationToolkitHarnessProject
Machine provisioning
Skill / agent definitions
Loop definitions
Loop instances + history
Swarm orchestration
Persistent knowledge
Engagement context packs
Local intent (AGENTS.md)

Derived from the repositories at agent-toolkit 4ea5878, agentic-workstation 1bbeb07, and agentic-harness a3f8822 (re-verified 2026-09-30).

Composable, not sequential

Which one do I need?

There is no required install order. Each project is a valid starting point, and each one works without the others.

  • I want portable AI capabilities without touching my machine setup

    Agent Toolkit

    One native binary, one install command. Deploys to Claude Code, Cursor, OpenCode, Copilot, Codex, Gemini CLI, Windsurf, Pi, and Muse Code — or as portable Agent Plugins bundles.

  • I want my whole machine reproducible

    Agentic Workstation

    chezmoi apply installs packages and the Toolkit; you adopt the rest at your own pace.

  • I want memory and context that survive across sessions and repos

    Agentic Harness

    Clone the workspace scaffold anywhere; it drives the Toolkit CLIs against your own state.

  • I want the full Personal DX workflow

    Agentic Harness

    Workstation provisions the machine, Toolkit executes, Harness remembers. Each part still works alone.