The State of AI Agents: Our Year in Review & What to Expect in 2027

2026 was the year of the AI agent. Or more specifically, the year of the AI agent coworker.

This past year, we saw agents move from experiments in the workplace to autonomous systems integrated with business applications that work alongside humans.

As a result, we also saw challenges emerge. It's not easy to manage dozens or hundreds of agents across different teams. And as they become more capable, how do you control what your agents can access and do? How do you keep them secure and accountable? Prevent them from blowing your budget?

These questions came up again and again as organizations grappled with agentic AI this year. In 2027, they'll become even more important.

Looking back at our blog this year, a few themes stand out. Here's what 2026 taught us, what to expect next year, and how organizations can prepare now.

What 2026 Taught Us About AI Agents

1. Agents need defined roles before they get autonomy.

Earlier this year, we explored how modern neoteams work. Humans and AI agents are now collaborating as teammates rather than humans simply using AI as a tool. Think of it this way: a tool does nothing until you pick it up. A teammate takes action on its own.

That's why AgentOps has become increasingly important. With agents acting on their own, you have to monitor and manage their behavior carefully to avoid anything going wrong. But before you can do that, you have to understand their role.

It helps to write job descriptions for your AI agents – basically treating them the same way as you do human employees. That means giving each agent a specific role, measurable expectations, clear boundaries, decision authority, and an escalation path.

2. Multi-agent systems break at the handoffs.

Dividing complex tasks among specialized agents is a great way to keep each one focused and optimized within its domain of expertise. The Achilles' heel is the handoff.

Context loss happens when one agent passes work to the next without including all of the information needed to execute the next step correctly. As the saying goes, a chain is only as strong as its weakest link.

We proposed that passing structured state is the solution. That is, carrying forward the goal, decisions made, open questions, and constraints from one agent to the next in a structured format.

Context loss in multi-agent systems can never be entirely eliminated, but its effects can be minimized.

3. Spending limits can't be left up to the agent.

We wrote a lot about agent budgeting this year: how to set practical spend limits before deployment, whether AI agents can manage their own budgets, and how to design cost-conscious workflows.

The common thread is that agents can optimize their spending, but infrastructure has to enforce the limits. These are architectural considerations. External controls are what allow the system to block or terminate actions in the moment through a real-time governance layer.

The most secure systems combine cost-efficient AI workflows with limits that the agent can't override.

4. Agents can create problems for people who never deployed them.

You might have your agents totally dialed in. Maybe they're running just the way you want them to, with costs under control and outputs as desired. But even then you may still have to deal with the mess left behind by other people's agents.

For instance, there's been a massive influx of agent-assisted pull requests that open-source maintainers have to spend time reviewing. The flood of AI slop creates extra work for people who never asked for it in the first place.

That's why we built Corbits Triage, which screens incoming pull requests against a project's contribution policy before a maintainer even looks at them.

But the broader lesson is that managing your own agents isn't enough nowadays. You also need to be prepared for the ones everyone else deploys.

What's Next for AI Agents in 2027

Agentic AI will grow faster than governance.

We asked earlier this year if AI agents are the new shadow IT. The answer is increasingly yes.

Gartner expects security or compliance incidents tied to unauthorized shadow AI to affect more than 40% of enterprises by 2030. Forrester predicts that, in 2027, one in three large Asia Pacific enterprises will uncover at least 10 employee-built shadow AI agents for every officially sanctioned agent.

And that's just the shadow AI problem. We're seeing an explosion of AI agents across organizations, whether they're officially authorized or not. By 2028, the average Fortune 500 enterprise is expected to have over 150,000 agents in use.

But only 13% of organizations believe they currently have the right AI agent governance in place... and the other 87% probably have good reason to be concerned they don't.

Agentic commerce will need proof of authorization.

Automated traffic already exceeds human traffic online. That includes ordinary bots as well as AI-driven automation; it is not a count of AI agents. As agents increasingly transact on our behalf, giving them access to a payment method remains inherently risky.

What is the agent allowed to purchase? How much money can it spend? How can a merchant or even your own team verify that a transaction was authorized, and by whom, for a defined limit and specific purpose?

Standards like Google's Agent Payments Protocol (AP2) use cryptographically signed mandates to define exactly that. Other developments in agent wallets and protocols are laying the foundation for growth in agentic commerce that could accelerate in the coming year.

Payment infrastructure of the future will have to ensure agent transactions are authorized, bounded, and auditable at all times.

Governance will be increasingly tiered instead of binary.

Gartner analyst Shiva Varma says the root cause of failure in AI agent governance is treating it as binary – either locked down or fully trusted.

If you apply the same controls to every agent indiscriminately, Varma explains, simple agents can become needlessly restricted, while more autonomous agents may receive too little protection. The result can be slower delivery, more shadow development, and greater operational, security, and compliance risk.

Agents that operate across different trust boundaries and at different autonomy levels shouldn't share the same rules. Instead, governance should be proportional.

Proportional governance is where you classify each agent by how much autonomy it has and match its controls to that level. With the right safeguards in place, you can avoid having to scale back or shut down production agents due to governance gaps.

So how do you prepare for a future with more agents and more autonomy... and more risk?

Getting Ready for 2027: A Five-Step Checklist for Agent Governance

If you're just getting started with agent governance, here are five practical steps to focus on first:

1. Inventory your agents. This includes employee-built agents and third-party tools. Remember: you can't govern what you can't see.

2. Write an agent job description. For each agent, define its purpose, identity and ownership, permissions, guardrails, decision authority, and escalation path. If you can't clearly describe an agent's job, you aren't ready to deploy it.

3. Classify your agents' autonomy level. Differentiate between agents that observe and advise versus those that act with human approval and those that act autonomously. Then you can match the controls to each distinct level.

4. Enforce limits outside the agent. Put guardrails like spending caps, access scopes, token and API key quotas, and rate limits in the infrastructure around the agent. This separates decision-making from enforcement.

5. Keep an auditable record of actions. Log what each agent did and when, what it accessed, under whose authority, and why. That's AgentOps in a nutshell.

Not sure where your agent governance stands? You don't have to figure it out alone. We're here to help.