Harness Architecture Comparison

Key dimensions

The pages in this wiki differ most clearly along five architectural axes: session container, memory substrate, work representation, evaluation loop, and deployment surface.

Comparison table

System	Session container	Memory substrate	Work representation	Evaluation style	Surface model
codex-cli	Threads / turns / items via codex-app-server plus cloud-delegated tasks	Session history plus repo docs and shared client state	Plans, docs, tool state, worktrees, automations, plugins	Self-review, GitHub auto-review, and enforced repo checks	CLI, IDE, web, app, cloud, and SDK/Slack control paths
claude-code	Fresh sessions with resumable artifacts, custom subagents, and experimental agent teams across separate sessions	CLAUDE.md, auto memory, feature lists, progress logs	Sprint contracts, subagents, agent teams, scheduled tasks, and explicit pass/fail features	Separate evaluator plus CI/review integrations and hooks	Terminal, IDE, desktop, browser, and remote-control surfaces
hermes-agent	Persistent multi-platform conversations and gateway-backed sessions	Searchable memory, skills, user modeling, API-backed reuse	Tasks, skills, cron jobs, profiles	Tool-driven verification and memory reuse	CLI, messaging, MCP, and OpenAI-compatible HTTP frontends
memento-skills	Persistent local sessions plus per-user IM sessions and stateful prompts	Structured markdown skills, local/vector/db skill stores, and layered runtime configuration	Retrieved skills, generated skills, reflection-driven rewrites, skill market downloads	Reflection loop plus static and execution-oriented skill verification	CLI, desktop GUI, local sandbox, and IM gateway surfaces
gas-town	Swarm sessions across named roles	Beads in Git / Dolt	Beads, epics, molecules, formulas, wisps	Human plus role-based oversight	tmux-style orchestrator/factory
gas-city	Modular orchestration nodes	Beads plus Wasteland federation	Builder primitives and wanted-board exchange	Federated trust and validator roles	Custom topologies over shared protocols
openclaw	Persistent service runtime	Workspace files, long-lived agent state, and layered skills	Embedded runtime plus ecosystem skills and integrations	Less explicit in current corpus	Cross-channel background service with a single main workspace

Main architectural lesson

The important divergence is not “which model is best” but where state lives and how work is represented. Codex externalizes protocol boundaries, Claude externalizes handoff artifacts, Hermes externalizes personal memory and skills, Memento-Skills externalizes learning itself into a writable skill library, and Gas Town externalizes the work graph.

Related pages

Read this page after agent-harness-anatomy and alongside orchestration-topologies, memory-persistence, and work-management-primitives. It is also the factual substrate for new-harness-design-notes.