SoulSpec (soulspec.org) is an open standard for AI agent identity. Its core claim: agent personality should be treated as source code — structured, versionable, shareable — not runtime configuration. Peter Steinberger, creator of OpenClaw (the fastest-growing repository in GitHub history with over 175,000 stars in 2025), pioneered the SOUL.md pattern, inspired by Anthropic’s constitutional AI work.
The file set:
| File | Purpose |
|---|---|
soul.json |
Package manifest: metadata, versioning, compatibility, discovery |
SOUL.md |
Personality: values, communication style, opinions, behavioral guidelines |
IDENTITY.md |
Who the agent is: name, role, backstory, contextual positioning |
AGENTS.md |
Operational workflows: task handling, tool usage, autonomous behaviors |
STYLE.md (optional) |
Communication patterns |
HEARTBEAT.md (optional) |
Autonomous check-in behavior — periodic execution configuration |
OpenClaw’s “Workspace-First” design philosophy treats these configuration files as the source of truth for agent identity and behavior. Multi-agent support is achieved through “one gateway with many brains” — each agent has isolated state, separate tools, different personalities. Not different models; different people.
Relationship to the AGENTS.md cross-tool standard: AGENTS.md emerged separately as a cross-tool standard for providing agent-specific project context. It originated in mid-2025 from collaboration between Sourcegraph, OpenAI, Google, Cursor, and others, and is now maintained by the Agentic AI Foundation under the Linux Foundation. As of March 2026, AGENTS.md is supported by Claude Code, Cursor, GitHub Copilot, Gemini CLI, Windsurf, Aider, Zed, Warp, and others. AGENTS.md provides project-level task instruction (what to build, what conventions to follow). SoulSpec’s SOUL.md provides agent-level behavioral specification (who the agent is, how it decides). These are complementary layers — the same distinction captured in Task Instruction and Role Specialization as Agent Configuration Layers.
Portability: Ollama 0.17 (February 2026) shipped with native OpenClaw integration, meaning local models can now consume SOUL.md files directly. The garden’s persona architecture is single-framework (Claude Code). As multi-framework agent deployment grows, the portability question will reach the garden.
The garden’s two-layer design (garden persona node + .claude/agents/ operational file) parallels SoulSpec’s structure. The garden persona node corresponds to SOUL.md + IDENTITY.md. The agent file corresponds to the operational deployment.
Three gaps in the garden model relative to SoulSpec:
soul.json, no mechanism for discovering, versioning, or sharing personas across projectsSoulSpec addresses portable identity across frameworks; it does not address how models mechanistically adopt personas (see [[The Persona Selection Model]]) or where persona traits live in activation space (see [[Persona Vectors and the Assistant Axis]]). Portability at the file level does not guarantee behavioral consistency across different model families.
The standard also does not address the coordination limits that apply when multiple specialized agents share a workspace. Research on multi-agent scaling found coordination gains plateau beyond approximately 4 agents — file-format portability does not solve this ceiling.