Agent Harness Capsules
The four concrete agent-runtime capsules expose Kevin's compiled knowledge about Claude Code, Hermes Agent, OpenClaw, and Codex as separate brain interfaces instead of one generic agent bucket.
Why This Exists
The wiki has many source pages about agent harnesses: architecture notes, tool pages, X bookmark ledgers, setup docs, project decisions, security warnings, and generated skill surfaces. Without capsules, agents load the generic Agent Harness page and miss the runtime-specific knowledge that makes each harness useful.
The capsule boundary is the fix. Each runtime gets a named public interface with:
- the canonical harness article
- the strongest supporting tool/bookmark/concept pages
- the skill routes that should be loaded first
- the tool and config surfaces that make the runtime operational
- maintenance rules for future ingests
This page is the human synthesis. The generated manifest lives in Brain Capsules and agent-docs/capsules.json.
Capsule Map
| Capsule | Load when | Center of gravity | Canonical sources |
|---|---|---|---|
claude-code-agent-stack |
Claude Code work, hooks, subagents, workflows, Claude skills, or Claude-to-other-harness transfer. | Terminal-first coding harness with skills, hooks, rules, memory architecture, and dynamic workflow fan-out. | Claude Code Harness, Everything Claude Code - Agent Harness Performance System, Claude Code Dynamic Workflows, Claude Code Memory Architecture (Reverse-Engineered), Claude Code Rules |
hermes-agent-stack |
Persistent personal agents, self-hosted runtime work, cron, memory, eval loops, profile distributions, /learn, memory/context providers, MoA, or Agent Machines Hermes deployment. |
Long-lived model-agnostic runtime with shareable profile distributions, skills as procedural memory, FTS/session memory, pluggable memory/context engines, cron, approvals, goals, multi-agent kanban, and model councils. | Hermes Harness (Nous Research), Hermes Agent (Nous Research), Agent Machines, The Eval Loop (Slop Is an Output Problem), Agent Memory System Architecture, Own the Learning Loop |
openclaw-agent-stack |
Multi-channel assistant gateway, OpenClaw skill/security work, messaging agents, ClawHub, Nodes, or OpenClaw as an Agent Machines runtime. | Local-first gateway daemon that connects messaging channels, nodes, skills, MCP/ACP, sandboxed sessions, and coding sub-CLIs. | OpenClaw Harness, OpenClaw Agent Skills, Agent Machines, Agent Guardrails Architecture, MCP Server Architecture, Security and Review Skills |
codex-agent-stack |
Codex app/CLI/cloud work, Codex automations, Record & Replay, OpenAI sandboxing, GPT-5.x-Codex, or source-compile scheduling. | OpenAI coding harness with Rust shared core, OS-level sandbox, AGENTS.md, MCP/App Server, cloud tasks, automations, and desktop skill capture. |
Codex Harness (OpenAI), GPT-5-Codex, Codex Record & Replay, Security and Review Skills, Scheduling Operating Model |
Decision Matrix
| Question | Prefer |
|---|---|
| Need a terminal coding agent with first-party Claude workflows, hooks, and subagents? | claude-code-agent-stack |
| Need a persistent self-owned agent that compounds memory and runs cron/eval loops? | hermes-agent-stack |
| Need a messaging gateway that can reach WhatsApp, Slack, Discord, Signal, iMessage, devices, and coding sub-CLIs? | openclaw-agent-stack |
| Need OpenAI's sandboxed coding agent, local/cloud/background Codex threads, or desktop workflow capture? | codex-agent-stack |
| Need the same skills and identity across all of them? | Cross-Agent Harness Portability plus the universal ~/.agents/skills bridge |
| Need to deploy one of them as a persistent worker? | Agent Machines |
| Need to compare them as products or architecture? | Agent Harness and the generated agent-harnesses.md pack |
Runtime Boundaries
Claude Code is strongest when the repo and shell are the center of the work. Its value comes from the harness around the CLI: hooks, file-aware rules, skills, subagents, dynamic workflows, memory compaction, and a growing public ecosystem around Claude-specific agent operations.
The Claude Code capsule should audit skill coverage against the nine-type portfolio in Agent Skill Bundles: reference, verification, data, team automation, scaffolding, review, CI/CD, runbooks, and infra ops. Product verification, code quality/review, runbooks, and infra ops are the proof-bearing layers; if they are missing, the capsule is only teaching the agent to produce work, not prove or operate it. Source: X/@om_patel5, 2026-03-18; Source: Anthropic Agent Skills docs, 2026-07-04
Hermes is strongest when the agent needs to live beyond the current terminal. Its durable primitives are memory, skills, cron, goals, approvals, messaging, model routing, and self-improvement loops. It is the default place to think about persistent agent behavior and eval-loop automation.
The Hermes capsule should load as a runtime stack, not as a single tool page. Read Hermes Harness (Nous Research) first for architecture, Hermes Agent (Nous Research) for current release/features, Agent Machines for persistent deployment, The Eval Loop (Slop Is an Output Problem) for improvement receipts, and Agent Memory System Architecture when the task touches retrieval, memory providers, or context engines. Profile distributions are the public package boundary: SOUL.md, config, skills, cron, MCP connections, and files can travel as versioned git state, while memories, sessions, API keys, and secrets stay local. /learn turns source material into procedural skill memory; it works best when the operator first stages a topic-specific Learnables room with source links, examples, pitfalls, and taste rules, then asks Hermes to codify the coherent thread. Memory provider and context engine plugins decide what the agent remembers or retrieves across sessions. MoA belongs in model-routing/council design, not in skill import. Source: X/@Teknium, 2026-05-31; Source: X/@jumperz, 2026-06-24; Source: Hermes profile distributions docs, 2026-06-30
Higgsfield Supercomputer is a hosted creative-agent product, but it belongs in the Hermes capsule as product-shape evidence: persistent agents expose role/employee packages, connectors, slash-triggered skills, project files, scheduled tasks, and memory as visible user primitives. Keep it as a reference for how agent capability is surfaced, not as a default local dependency. Source: X/@higgsfield, 2026-05-13; Source: Higgsfield Supercomputer product page, 2026-06-30
OpenClaw is strongest when the agent is a gateway across channels and devices. It is less "coding assistant in a repo" and more "local-first assistant daemon with skills, sessions, Nodes, MCP, and message routing." Treat third-party skills as supply-chain risk.
The OpenClaw capsule should load OpenClaw Agent Skills when the task touches shared workflows: /autoreview, /handoff, Crabbox/Testbox, transcript provenance, or session-viewer tooling. Local decision: autoreview and handoff are already active; crabbox is documented; agent-transcript and session-viewer are import candidates only after privacy and redaction review.
Ahmad's "switch from OpenClaw to Hermes" bookmark is a capsule-routing signal, not a merge instruction. Keep Hermes and OpenClaw as separate interfaces: Hermes owns the persistent self-owned runtime lane, while OpenClaw owns the gateway/channel/node lane. Source: X/@TheAhmadOsman, 2026-06-22
Gemini Spark belongs near this map as a comparison point, not as a Kevin-owned capsule. It demonstrates the same user-visible primitives - background execution, tasks, skills, schedules, connected apps, and consent checks - inside Google's hosted Gemini control plane. Use it when comparing consumer managed agents against OpenClaw/Hermes, but do not route repo work or personal wiki runtime decisions through it. Source: Gemini Spark overview, 2026-07-04
Codex is strongest when OS-level sandboxing, OpenAI model/runtime integration, background tasks, automations, and desktop Record & Replay matter. Its local contract is AGENTS.md, ~/.codex/skills, sandbox policy, MCP, and the Codex thread/turn model.
Shared Invariants
All four capsules inherit the same Kevin rules:
skills/.runtime/allis the repo-generated skill projection.~/.agents/skillsis the universal home bridge.- Bespoke harness skill directories should point or symlink to
~/.agents/skills. - Identity and operating rules come from the wiki, not per-runtime copies.
- New bookmarks or docs should update the runtime capsule source pages before packs regenerate.
- Version snapshots are required for promoted tools, SDKs, CLIs, and harness features.
- Fan-out is only a harness improvement when it has independent verifiers and retry loops; otherwise it is just parallel slop generation.
- Operators should watch blast radius and interrupt/status-check when observed scope diverges from expected scope.
Failure Modes
| Failure | Fix |
|---|---|
| Runtime pages drift into isolated notes. | Add the new source to the correct capsule and regenerate packs. |
| A Claude/Codex/OpenClaw/Hermes skill gets copied into a bespoke folder. | Move or link through the canonical skill source tree and ~/.agents/skills. |
| A bookmark is represented only in an ingest ledger. | Promote the durable idea into the runtime page or this capsule map, then update the artifact audit row. |
| A harness is treated as just a model. | Route through Agent Harness: model, runtime, tools, memory, permissions, verification, state, and UI are separate layers. |
| A persistent agent lacks proof. | Add evals, cron receipts, state logs, approval trails, or run artifacts before calling it operational. |
/learn is treated as random memory dumping. |
Stage a topic-specific learning room first, add sources/examples/pitfalls/taste rules, then codify the coherent thread as a skill or reusable artifact. |
/learn is treated as enough for repeated retrieval. |
Keep procedural instructions in skills, but move constantly reused knowledge into memory providers, context engines, wiki pages, or capsule load rules. |
Timeline
- 2026-07-04 | Added the Claude Code skill portfolio audit to the capsule boundary so generated packs inherit the rule: a harness capsule needs reference, verification, data, team process, scaffolding, review, CI/CD, runbook, and infra-ops coverage. Source: X/@om_patel5, 2026-03-18; Source: Anthropic Agent Skills docs, 2026-07-04
- 2026-06-30 | Added shared fan-out and operator-control invariants from the
/orchestrateand Peter Steinberger bookmark review: planners/workers need verifiers, and long-running agent work should be stopped or split when blast radius grows unexpectedly. Source: X/@cursor_ai, 2026-05-07; Source: Peter Steinberger, read 2026-06-30 - 2026-07-04 | Added Gemini Spark as a hosted personal-agent comparison point while keeping Kevin-owned harness capsules limited to Claude Code, Hermes, OpenClaw, and Codex. Source: Gemini Spark overview, 2026-07-04
- 2026-07-03 | Added Ahmad's OpenClaw/Hermes routing bookmark as a capsule-boundary guard: Hermes and OpenClaw remain separate runtime interfaces even when Hermes is preferred for persistent self-owned agents. Source: X/@TheAhmadOsman, 2026-06-22
- 2026-06-30 | Added Hermes profile distributions to the capsule boundary: a capsule can expose public brain/config/skills/cron/MCP as versioned git state, while private memory and secrets remain outside the package. Source: X/@Teknium, 2026-05-31; Source: Hermes profile distributions docs, 2026-06-30
- 2026-06-30 | Added Higgsfield Supercomputer to the Hermes capsule as product-shape evidence for exposing skills, files, schedules, connectors, and memory as user-visible agent primitives. Source: X bookmark artifact audit, 2026-06-30
- 2026-06-30 | Attached OpenClaw Agent Skills to the OpenClaw capsule after reviewing
openclaw/agent-skills;autoreviewandhandoffstay active, transcript/session tools remain candidates pending privacy audit. Source: X/@hungv47, 2026-06-05 - 2026-06-30 | Deep-reviewed the Hermes Agent bookmark cluster and updated
hermes-agent-stackto include the v0.17 Reach release,/learn, memory provider and context engine plugins, MoA, and skill-candidate boundaries. Source: X/@NousResearch, 2026-06-19; X/@Teknium, 2026-06-23 and 2026-06-26; X/@mr_r0b0t, 2026-06-26; Hermes docs - 2026-06-29 | Created after Kevin requested full capsules for Claude Code agents, Hermes agents, OpenClaw agents, and Codex, with docs, bookmarks, concepts, and tool pages aggregated into separate public brain interfaces. Source: User request, 2026-06-29