HUMΛN
Architecture
Architecture

Agents Before and After a Trust Layer

HUMΛN Team··8 min·Technical

“Reschedule my meetings tomorrow, update my project board, and let my team know.”

In both worlds, the user asks in chat. The agent talks to calendars, the PM tool, and chat. The difference is everything between the agent and those systems. For the same “reschedule my life” use case, an OpenClaw-style agent and a HUMΛN-governed agent look similar on the surface—but completely different when you trace identity, policy, and provenance.

OpenClaw-Style Flow

User → chat channel → OpenClaw gateway → LLM → skills (shell, browser, API keys/config). Skills talk directly to Google Calendar, Jira/Linear, Slack/Teams. The agent runs with the same rights as the host OS user. Config files hold secrets. There’s no central policy or identity beyond “this instance.” Fast to ship. Terrifying at security review.

HUMΛN-Governed Flow

User → channel (Slack, WhatsApp, etc.) → Companion. Companion → HumanOS, which: checks Passport (who is this human, which org?), consults policies (what’s allowed, what needs confirmation?), uses the Capability Graph (which connectors, what scopes?). HumanOS → connectors with scoped, Passport-bound tokens: “read/write calendar,” “update project board,” “send team notification.” Every action is logged with identity, capability, and extension ID. Same high-level UX. Very different semantics: who is allowed to do what, and how we prove it later.

Side-by-Side

Dimension Typical OpenClaw-style runtime HUMΛN-governed runtime
Identity Instance / host user Passport + delegation; every action attributable
Policy Prompts, discipline HumanOS; first-class allow/deny/confirm
Secrets Config files, env, API keys Passport-bound, scoped tokens; no raw keys
Scope OS-level (whatever the user can do) Capability-level (declared, connector-backed)
Auditability Logs if you added them Full chain: who, what, when, under what delegation
Blast radius Compromise = full user/OS Compromise = only delegated capabilities

Why This Matters for Builders and Buyers

Builders: With an OpenClaw-style design, you ship fast and hit a wall at security review. With a trust layer, you can still ship fast—but in a way that can pass a questionnaire. Same power; different blast radius.

Buyers: You have to look at how an agent interacts with systems, not just which features it demos. “Can it reschedule my calendar?” is the wrong question. “How does it prove who did what, and who can revoke it?” is the right one.

Closing

Agents are here. The difference between “we banned it” and “we bet the company on it” is the presence of a trust layer. We’re not fighting the wave of agent runtimes. We’re building the substrate they’ll all eventually need to plug into.


For more on why this matters and how we define HUMΛN-compliant extensions, see OpenClaw and the Cost of Skipping the Trust Layer and What a “HUMΛN-Compliant Extension” Actually Means.