Auth is Just the First Step

Authentication gets AI agents connected, but secure production access also requires authorization, permissions, governance, and continuous control.

AI AGENTS

August 28, 2026

Kushal Chatterjee

Summarize with AI:

Many companies are now touting the ability to provide authorizations for AI tools, making data integration for agents “simple”.

Our view is that managed auth is the bare minimum for companies building complex workflows, especially with agents. Everyone wants to be able to hook up their Claude or ChatGPT to a variety of business tools, and have their credentials managed. If you don’t provide this as part of your integration service, your product is dead on arrival.

But auth is just the first step for agentic workflows. It solves one fundamental issue, but it doesn’t suddenly make agents production ready.

If your goal is to build real agents that can be deployed in enterprise systems, auth is not enough.

Where auth only tooling falls short

What happens after you’ve authorized all of your integrations?

You start dealing with an entirely new category of problems, such as:

Memory. An agent has to remember when it learned on its previous iteration. Just having credentials doesn’t ensure knowledge is retained or organized in a manner that facilitates recall.

Discovery. Agents must be able to search efficiently and accurately across disparate business systems. This is incredibly complex, even with MCP servers setup for each tool. Without control or indexing of the data, agents must query sources individually and paginate for results.

Entity resolution. An agent has to be able to understand that Sarah Chen in Salesforce is also S. Chen in a Linear ticket, while understanding that she is sarah_c@datasystems.co in Hubspot. Otherwise, it won’t have the full context when making decisions that affect the entity in question, which is oftentimes a customer we want to keep happy.

What do you notice here? Auth doesn’t solve any of these problems.

Why you should be concerned

So, those are the unsolved problems.

For many people using AI on a daily basis, the response could be a shrug. They can work around it, or their workflows are for simpler, personal tasks, and these problems aren’t big enough.

But if you’re building an agent that will affect an entire organization, the consequences can be catastrophic. If you rely on auth for production agents, these are some of the fun little nightmares you can expect to deal with:

The memory gap. Every agent invocation starts from zero: re-fetching the same contacts, re-paginating through the same ticket history, re-discovering context it already had five minutes ago. The auth tooling authenticated your agent to Salesforce. It did not remember what the agent learned there. So you build a persistence layer yourself. You pick your vector database of choice and manage the pipelines and indexing. You wire up something like Hyperspell or Mem0 or a hand-rolled Postgres setup to give your agent a memory it should have had from the start.

The token tax. This is a direct result of the discovery problem I explained earlier. Pagination is extremely costly. When you pay for a Claude or ChatGPT plan, you want your tokens to be spent on reasoning, not fetch and discovery. With most of the costs being focused there, it simply isn’t possible to build a production agent that will be efficient at scale.

The governance vacuum. Auth is just a tiny part of governance, which is credential and API key management. This is fine when one person is using the agent. But at production scale, with multiple teams running agents against shared business systems, ungoverned write access is the thing that gets someone fired. You need per-member, per-entity, per-action permissions configured in one place, not scattered across application code in six different repos.

To be clear, we don’t feel that auth tools are ineffective. They solve one very specific and important part of the problem, and for many smaller use cases, that could be sufficient.

At production scale, it simply isn’t.

Here’s how we build beyond auth

Airbyte Agents comes with OAuth, but it also builds beyond that to deliver a foundation for real production-grade agents.

We provide the Context Store, a search optimized, replicated index of your data that is assembled before runtime. This solves the discovery problem of agents probing multiple APIs and MCP servers to put together context and parse through it for the right answer. Our benchmarks indicate that the agents powered by the Context Store consume 80% fewer tokens and 40% fewer tool calls, resulting in up to 90% lower costs on multi-source queries.

We provide cross-source entity resolution and querying to provide agents full context when executing actions. Airbyte’s agent connectors all possess direct read capabilities, and many are capable of writing directly to systems. But beyond handling auth, they also come with governance policies. Admins are able to separate users into workspaces with isolated credentials, and set entity policies within those credentials for different users.

Critically, every single action an agent makes with Airbyte is logged and auditable.

That’s what real agentic data infrastructure looks like in practice. If you’re interested to learn more, read the docs, or try Airbyte Agents for free.