The security industry just priced the agent problem, and the number is $100 million. On July 20, 2026, a startup called Neo exited stealth with a round led by Andreessen Horowitz and Bessemer Venture Partners (Craft Ventures and Merlin Ventures participating), founded by veterans of SentinelOne, Wiz, and Palo Alto Networks — including former SentinelOne COO Nick Warner (The Next Web; GlobeNewswire).
Here's the part that matters even if you'll never be Neo's customer: the thesis is copyable, and you should copy it this week.
What Neo is selling#
Neo names the category "Agentic Software Control." Stripped of the pitch deck, the product is a real-time control layer that does three things (MSSP Alert):
- Inventory — discovers every AI agent, model, extension, and MCP server running inside a company.
- Policy at the endpoint — enforces fine-grained rules over tool calls, API access, and data movement.
- Audit — keeps an immutable trail that traces every action back to the human or agent that originated it.
The bet behind the round is a single Gartner number: only about 5% of enterprise applications had agentic capabilities in 2025, and roughly 40% will by the end of 2026. In one year, the count of autonomous, credentialed actors inside a business multiplies — and the tools built to watch humans and static software weren't built to watch things that invoke tools and move through workflows on their own.
Why a team of one has the same problem#
It's tempting to file this under "enterprise" and move on. Don't. The exposure isn't proportional to headcount.
If you're two people running Claude Code, a couple of browser agents, a nightly cron that calls an LLM, and three MCP servers, you already have autonomous processes holding real credentials — your production database, your Stripe key, your email account. That's the same blast radius Neo sells its platform to contain, minus the security team and minus the $100M to build one. The non-human-identity attack surface we mapped for founders is exactly this, and it's why agent identity itself is now a funded category.
You can't buy Neo's answer at your stage. You can copy its shape.
The cheap version of agentic software control#
You don't need a control plane. You need the three primitives Neo raised $100M to package, implemented by hand:
- Inventory. Write down every agent that holds a credential and what it can reach. If you can't list them, you can't secure them — start with our agent inventory walkthrough before a breach forces the exercise.
- Policy at the endpoint. Kill the shared master key. Give each agent a scoped, short-lived token so a compromised one can touch only its slice — the scoped-credential pattern is a weekend of work, not a quarter.
- Audit. Log every tool call with which agent and which human triggered it. Even a JSONL file beats nothing: when something surprising happens, you want it to trace back to exactly one identity you can revoke.
That's the same inventory → policy → audit spine the enterprise version runs on — and it's the direction the whole agent-runtime-governance category is consolidating around, from the GA agentic-security launches this week to the economics of AI-native attacks.
The tell#
When a16z and Bessemer put $100M behind ex-SentinelOne operators to inventory and govern agents, they're not betting on a niche. They're betting that within a year, every company — including yours — will be running software it can't fully account for. The founders who come out ahead aren't the ones who buy the control layer. They're the ones who never let the sprawl start: one key per agent, one log for every action, one human answerable for each. Do the cheap version now, and you'll never need the expensive one.
