How We Built Truly Autonomous Agents

What if your AI could simultaneously analyze 50 contracts, research market conditions, update your CRM, and create a board-ready presentation – all while you grab coffee? We've spent the last year turning this vision into reality. Here's what we learned.

Evolving from Programmatic to Autonomous

We've been deploying agentic workflows for over a year now, allowing customers to build guided, adaptive processes that dynamically respond to AI outputs. These programmatic agents already revolutionized how we handle document preparation – automatically processing, categorizing, transforming, and embedding information in ways that were never before possible.

But we wanted to go further.

Enter the Autonomous Agent

Autonomous agents take the dynamic nature of agentic workflows an order of magnitude further. Instead of following pre-defined paths with AI-powered decisions, they put the model fully in control – reasoning, planning, and even working with various tools to deliver requested outcomes.

Here's a secret that might disappoint (or delight) you: the core of their reasoning is just a loop, enabling the agent to recursively and iteratively execute thoughts and actions.

That's it. The model is run in a loop, selecting tools, executing them, reviewing its results, and repeating until it achieves its goal. As it turns out, intelligence is just a while loop.

So the core is simple, but as with everything in enterprise software, the devil is in the details, and in the scalability of the basic principles..

The Power Is in the Tools

As we discussed in our previous blog, what determines an autonomous agent's potential isn't the while loop – it's the tools that it has at its disposal. Just as RAG quality depends on context, agent capability depends on tooling and how these various tools can be combined to solve various business challenges.

Our tool ecosystem ranges from foundational to highly sophisticated:

Foundational Tools

• Think. Deep analysis and problem decomposition.
• Plan and Update Plan. Create and track structured, executable plans with visual updates on progress.
• Search Documents. Intelligent document discovery with context-aware filtering.

Advanced Capabilities

• Execute Parallel Work Streams. This is a game-changer. Decomposes complex problems into parallel sub-agents, each working independently.
• Analyze Spreadsheet. Executes custom code to analyze Excel spreadsheets. 
• Update  Document. Update a document — including Microsoft Word docs — using a series of patches, without rewriting the entire document. This is critical to being able to track changes and effectively handle long-form documents (more than a couple of pages).
• Collection Management. A “collection” is a working structure of documents for the agent. Collection Management allows us to create collections, update collections, and even add documents to collections to better organize insights.

The above capabilities are just some examples. View the full list here.

A Real-World Example

To tie this all together, let’s take a look at how an agent handles a complex enterprise task, in this case an analysis of contract renewals:

Now imagine if the agent had hundreds or even thousands of contracts to analyze. How long would that take an agent? Maybe hours. How long would that take a team of people? Days, perhaps even weeks. This is the immense power of autonomous agents.

What’s Really Going on Behind the Scenes?

At Vertesia, we like to say that AI and agents may feel a bit like magic, but the real magic is how all the little things come together to make your agents work. It’s a bit like landing on the moon.

So what’s really going on behind the scenes and what are some of the really hard problems we solved here?

Working Memory Overload

Context windows feel infinite for single queries but become cramped during complex, multi-step reasoning. Imagine analyzing federal tariff rulings against your entire product catalog – that's gigabytes of context.

Our Approach:

• Agent-driven content delegation. When an agent needs to analyze a document, it doesn't load all of the content. Instead, it delegates to a specialized sub-agent with a focused prompt. The main agent receives only the analysis.
  
  
• Dynamic checkpointing. When memory approaches limits, the system automatically compresses the working memory into a detailed summary, allowing the agent to continue without losing context.
  
  
• Distributed persistence. Using Temporal as our orchestration layer, agent state persists across hours, days, or even weeks. System crashes, deployments, or network failures don't matter – agents resume exactly where they left off.

Security & Permissions

Enterprise agents can't operate with root access. They must respect existing permissioning while maintaining flexibility.

Our Approach:

• User-inherited permissions. Agents receive single-use tokens inheriting user permissions (or less).
• Short-lived, refreshable tokens. Tokens are automatically refreshed based on system or user actions.
• Context-aware tool scoping. When reviewing a specific object, search tools automatically scope to that object's data zone.

Dynamic Tool Adaptation

Tools can’t be static – they have to adapt to the context of the work to be completed:

• Reviewing insurance claims? Search automatically scopes to the claim's data.
• Need version control? update_document forces versioning for every change.
• Searching databases? Tool descriptions dynamically include your schemas and properties.

Agent Swarms

The real magic happens when agents work together. Using execute_parallel_work_streams, a master agent can orchestrate entire teams of specialized agents, each contributing their expertise to solve complex, multi-faceted problems.

Lessons Learned

1. Start with the tools. Agent intelligence is bounded by tool quality.
2. Memory management is critical. Elegant solutions beat brute force.
3. Security can't be an afterthought. Build permissions into the architecture.
4. Persistence enables real work. Long-running agents need rock-solid state management.
5. Parallel execution changes everything. Complex tasks become tractable when decomposed.

What's Next?

We're seeing autonomous agents transform how enterprises handle:

• Contract analysis and negotiation
• Compliance monitoring and reporting
• Customer intelligence aggregation
• Strategic planning and forecasting

The question isn't whether autonomous agents will transform enterprise work – it's how quickly organizations will adapt to this new paradigm.

Want to see autonomous agents in action? Schedule a demo to see how they can transform your business.

At Vertesia, we're building the infrastructure for enterprise AI agents. Our platform handles the complex challenges of memory, persistence, security, and tool orchestration so you can focus on solving business problems.