By Jon Cheney, founder of GenAIPI. Published August 3, 2026.
I want to start by saying something nice about the competition: the current generation of AI agents is genuinely remarkable. OpenClaw crossed 145,000 GitHub stars because it works. Hermes made a self-improving agent real enough that people run their businesses' daily grind through it. Viktor built an "AI employee" that lives in Slack and touches thousands of tools. And ChatGPT and Claude remain the best thinking partners most professionals have ever had.
And yet, week after week, companies come to us after trying all of the above — because none of it stuck. Not because the AI wasn't smart enough. Because the AI didn't know who was asking, wasn't allowed to know what it could touch, and had no one in the building responsible for making it actually work.
The pattern: individual tools, team problems
Every tool above was designed, first, for an individual. One person, one context, one set of credentials. That design assumption is invisible when you're testing an agent on your own laptop — and fatal the moment you deploy it across a team.
A team is not a big individual. A team has an org chart. It has a finance department whose numbers the sales team shouldn't see, HR records that almost no one should see, contracts under NDA, and a new hire who joined on Monday with different access than the veteran she sits next to. Every serious business system of the last thirty years — email, file storage, ERP, HR — was built around this reality.
Most AI agents simply weren't.
OpenClaw: power without guardrails
OpenClaw is the most capable open-source agent framework available, and I understand why developers love it. But the security record this year tells the story for teams: dozens of publicly disclosed vulnerabilities (CVEs), a chained exploit path (the "Claw Chain," described in published Cloud Security Alliance research) that researchers reported could enable full agent takeover, and security-vendor advisories warning of reported exploitation attempts inside enterprise networks. Trade publications now run CISO checklists specifically about open agent-framework risk. (These figures reflect public reporting as of mid-2026 — check current disclosures when you evaluate.)
None of this means OpenClaw is bad software. It means it was built for power users first, and the governance layer — who can invoke it, what it can reach, what it must never do on whose behalf — was left as an exercise for the adopter. Most companies never do that exercise. Security researchers have started calling this the agentic AI governance gap, and OpenClaw deployments are where the gap shows up first.
Hermes: a brilliant operator — for one person
Hermes, from Nous Research, is genuinely innovative: an agent that creates skills from experience and builds a model of you across sessions. Read the reviews, though, and notice the pronoun. You. Hermes is a personal operator. It's self-hosted on infrastructure you own, it learns one person's patterns, and its security model is essentially "it has whatever access the machine it runs on has."
That's fine for a founder running solo. For a team, a learning loop with no permission awareness is a liability that compounds: the agent gets better and better at doing things no one decided it should be allowed to do.
"AI employees": the right metaphor, half-implemented
Products like Viktor deserve credit for getting the metaphor right — an AI employee should have the access of an employee, not the access of an admin. Their model bolts membership onto each connected tool: invite the AI to the channels and tools it should see, like onboarding a contractor.
It's a real step forward. But bolt-on membership per tool is not the same as understanding your organization. It can't answer "who in Finance can see payroll?" It doesn't reason about roles, departments, sensitivity levels, or the difference between what an intern and a VP should get back from the same question. And when the tool count grows past a handful, per-tool membership management becomes its own unmanaged sprawl.
ChatGPT and Claude: assistants, not an operating layer
The frontier assistants are extraordinary at what they are: thinking partners. But a chat window has no memory of your org structure, no connection to your systems of record, and no concept of execution with approvals. Teams that standardize on "everyone gets a chatbot license" get exactly that — a thousand disconnected conversations, copy-pasted into the real tools by hand. That's an accessory to work, not an operating layer for it.
What teams actually need
After hundreds of company engagements, we've concluded the requirements are stable across industries:
- Permission-aware intelligence. The AI must know who is asking and answer within that person's actual access — combining role-based access control (RBAC: what your role entitles you to) with attribute-based access control (ABAC: context like department, sensitivity, project, time). Microsoft, Okta, and WorkOS have all published "least privilege for AI agents" guidance this year. This is becoming the standard, and most agents don't meet it.
- Visibility. You can't govern what you can't see. Leaders need a live map of who — human or AI — can access what across the company. (This is exactly why we built Atlas View into G.E.N.A.)
- Autonomy with accountability. Agents should execute real work — but through approval flows, with an audit trail, as a governed member of the org rather than a root-access daemon.
- Human leadership. This is the part no vendor wants to say: even the best tool fails without someone responsible for it. Adoption, training, workflow redesign, governance — that's leadership work. It's why our fractional Chief AI Officer teams install and run G.E.N.A. rather than tossing a login over the wall.
The other side of AI
We built G.E.N.A. — the Generative Execution Neural Architecture — because we kept meeting companies that had tried the tools above and ended up with either a security incident, a single power user's toy, or shelfware. G.E.N.A. is permission-aware at its core (RBAC + ABAC combined), ships with Atlas View for live access visibility, executes autonomously within governed boundaries, and arrives with a human fCAIO team that makes it real inside your company.
Over the next three days I'll publish detailed comparisons: G.E.N.A. vs. OpenClaw, G.E.N.A. vs. Hermes, and G.E.N.A. vs. Viktor. If you'd rather just see it, schedule a strategy call — we make AI work for you.