Arcade, Conduct, and Optiak are a sign that the agent market is maturing fast: the money is moving away from “can it do the task?” and toward “who controls it, who can trust it, and how does it fit into the enterprise stack?” The next winners won’t just ship clever agents; they’ll own the rails around execution, permissions, and reliability.
The demo era is ending
For the last year, “agent” often meant a polished demo: a chatbot that could click around a browser, draft an email, or chain a few tools together. That phase produced plenty of excitement, but it also exposed the same three failure modes over and over:
- agents do the wrong thing too confidently
- enterprises can’t safely give them access
- teams can’t measure whether they’re actually working
The recent funding wave shows investors now care less about novelty and more about operational control. That’s why Arcade’s $60M Series A, Conduct’s $60M Series A, and Optiak’s €4M pre-seed matter. They point to three different layers of the stack that are becoming strategic: authorization, orchestration, and integration.
The real battleground: control, not cleverness
Arcade: the permission layer for agents
Arcade.dev’s $60M Series A is one of the clearest signals in the market. Arcade is building an authorization and security platform for AI agents, focused on managing what agents are allowed to do in production systems.
That sounds unsexy until you think about what enterprise agents actually need:
- access to customer data
- permission to create tickets, issue refunds, or trigger workflows
- guardrails around sensitive actions
- auditability when something goes wrong
If an agent can reason but not act, it is a toy. If it can act without fine-grained permissions, it is a liability. Arcade is betting that the valuable layer is the one between the model and the system of record.
That’s a strong wedge because it is infrastructure, not application logic. As more companies deploy multiple agents across functions, someone has to enforce policy consistently. If Arcade becomes the default control plane, every agent in the company passes through it.
Conduct: owning execution inside the enterprise
Conduct’s $60M Series A is a different bet, but the logic is similar. Rather than selling “an agent,” Conduct is building the environment in which agents actually operate. In practice, that means the winner is the company that makes the agent useful across workflows, not just impressive in a sandbox.
This is where enterprise integration becomes the moat.
An agent platform that can:
- connect to existing tools
- route tasks across systems
- preserve context across steps
- keep humans in the loop where needed
…is much harder to rip out than a standalone chatbot. Conduct is well positioned if it can become the layer where execution happens reliably enough for real business processes.
That is the market shifting from one-off automations to production systems. Buyers no longer ask, “Can it do this once?” They ask, “Can it do this every day, under policy, with logs, approvals, and fallback paths?”
Optiak: the orchestration layer in the middle
Optiak’s €4M pre-seed is smaller, but strategically important. Optiak develops a modular operating system and orchestration layer that sits between enterprise applications and underlying AI models.
That framing is telling. The most durable agent companies may not be model companies at all. They may be middleware companies that decide:
- which model handles which task
- when to invoke tools
- how workflows should branch
- how context moves between apps and agents
If Arcade is about who is allowed to act, and Conduct is about where action happens, Optiak is about how action gets coordinated. In an agent stack, orchestration is not plumbing. It is product definition.
The startups that control this layer get leverage in two directions:
- downward, because they abstract over model providers
- upward, because they become embedded in business workflows
That’s why even a modest round like €4M can matter. If the product becomes the operating layer, it can turn into a strategic choke point.
Evals and observability are becoming table stakes
The other major signal in the market is that reliability tooling is getting funded as a category, not as a feature. Probably’s $9M Seed positions deterministic validation as a serious product category, while ChatSee.ai’s $6.5M and ChatSee’s $6.5M Seed both point to a growing appetite for failure intelligence and observability in autonomous systems.
That makes sense. Enterprises will not scale agents based on vibes. They need to know:
- where the agent failed
- why it failed
- whether the output was valid
- how often it deviated from policy
- what changed after a model update
This is the same pattern that played out in cloud, DevOps, and security: the first wave of products proves the technology works; the second wave proves it can be trusted.
The implication is brutal for everyone else. If your agent product does not own execution, permissions, or evaluation, you may still have a business — but you may not have a category.
Why some companies become platforms and others become features
The market is splitting into two camps.
The winners
These companies do one of three things well:
- Own the action layer
Arcade manages what agents can do.
- Own the workflow layer
Conduct and Optiak help agents operate inside real enterprise environments.
- Own the trust layer
Probably and ChatSee help enterprises validate, monitor, and debug agent behavior.
These are durable because they sit close to the risk center. In enterprise AI, risk is the product.
Everyone else
The rest are increasingly becoming features:
- an agent inside a vertical SaaS app
- a point solution for a single workflow
- a demo that gets replicated by a larger platform
- a model wrapper without control or trust
That doesn’t mean those startups won’t raise money. It means the bar for defensibility is rising. If the product can be copied by a platform vendor, procurement software suite, or major cloud provider, investors will eventually price that risk in.
The funding map says the same thing
A few other recent rounds reinforce the pattern:
- NeuralTrust raised $20M Seed to build AI security and governance tools for enterprises building generative AI and agentic systems.
- Pints AI raised $5.6M Pre-Series A with an auditable platform for regulated financial institutions.
- ENT raised $100M Seed for intent-aware endpoint security that evaluates human and AI agent behavior in real time.
- CNTXT AI raised $60M Series A around sovereign AI infrastructure and enterprise data solutions.
- Undo raised €31M Series B with deterministic program recording technology for AI coding agents and developers.
Different categories, same direction: the market is rewarding companies that make agentic systems safer, observable, and easier to deploy in enterprise environments.
What to watch next
If you want to know who is winning in agents, ignore the flashy demos and look for these signals:
- they are attached to permissions, policy, or audit logs
- they integrate deeply with enterprise systems
- they make failures measurable
- they reduce human review without eliminating human control
- they become the layer other agents depend on
That is the real shift. The agent market is not just about intelligence anymore. It is about who gets to decide, who gets to act, and who gets to prove it worked.
Takeaway for sellers to AI startups
If you sell into AI startups, stop pitching “growth” as a generic category. Sell to the pain around control, reliability, and integration: security reviews, eval workflows, auditability, orchestration, and deployment plumbing. That’s where budgets are forming, and that’s where the next generation of agent winners is building.