Inside Field Orchestrator, our multi-agent harness for audit

From tasks to workflows

I still fondly remember the days of building AI products with GPT 3.5. The context window was only 16,385 tokens and while it felt so powerful back then, in reality we could only automate single tasks.

Fast forward to 2026 and we live in a totally different reality. Today, coding agents can perform software tasks that would take humans many hours and task length roughly doubles every seven months.

At Fieldguide, we’re building agents that can complete full audit workflows. These workflows can span hundreds of documents, run for hours, and require auditor input and review along the way.

To support this, we built Field Orchestrator, Fieldguide’s multi-agent harness and the top-level agent that runs on it. It lets us decompose complex audit workflows into units of work that run in parallel while preserving shared context, following a firm's methodology, and keeping the auditor in the loop.

Why a multi-agent architecture?

Auditors specialize in developing a deep understanding of companies and industries. They understand the client’s business, the firm’s methodology, and they apply their professional judgment to assess risk and reach conclusions.

Supporting those judgments takes significant effort. In a typical testing workflow, auditors analyze large population files, apply sampling methods, and collect supporting evidence. They then search through those documents, link evidence to sampled items, extract values, and compare information across files. Throughout the process, auditors exercise professional judgment to reach a conclusion.

AI can provide significant value by reducing the time it takes to complete this analysis and by finding the relevant evidence, allowing auditors to spend more time investigating exceptions and applying judgment.

For agents to perform end-to-end workflows, a multi-agent architecture enables us to overcome several challenges:

1. A single test can easily involve hundreds of supporting documents, which quickly exceeds the reliable context and attention of a single agent.
2. Different steps in the workflow require distinct context and specialized tools.
3. Individual samples are independent and can be tested in parallel.
4. Field Orchestrator needs to retain high-level context while interacting with the auditor and coordinating specialized agents. It doesn’t need to manage every operation—just the high-level details.

How Field Orchestrator works

Auditors need the ability to review intermediate outputs, provide additional context, and at times intervene. To support this, Field Orchestrator runs on a live WebSocket server. This allows the auditors to interact with Orchestrator and enables us to emit status updates, request additional context from the auditor, and have them provide their judgment.

The Orchestrator agent harness supports multiple AI workflows throughout the Fieldguide product, but we will specifically focus on testing in this post. When Orchestrator starts, it gathers the relevant context and receives a guidance document that describes the firm’s testing methodology. It uses these inputs to build a plan that the auditor can then refine and approve before executing.

Once the auditor approves the plan, Orchestrator coordinates the execution across a set of sub-agents to perform the work.

It delegates population analysis and sampling to our spreadsheet agent. The spreadsheet agent inspects, writes, and executes code to read and modify the sheet. After it completes its analysis and applies the sampling strategy, it returns an updated spreadsheet showing the work performed and generates a sample sheet. It also sends Orchestrator a detailed summary of the spreadsheet’s contents, along with the analysis and sampling performed.

The sample sheet becomes the foundation for the next stage: collecting evidence, linking documents to sample items, extracting values, and evaluating them against the testing procedure. Orchestrator stays focused on executing the plan and interacting with the auditor, while sub-agents focus on their tasks and produce concrete outputs such as spreadsheets, extracted values, citations, and reasoning that auditors can review and correct.

The road ahead

Building AI product has been the most enjoyable challenge of my career. Working alongside auditors to understand how they make decisions, analyze evidence, and review work has shaped everything, from the tools we give agents to how we design and build the UI.

The shift from a system of record to a system of work is a once-in-a-lifetime moment. Even with this year’s advances in agents, we’re still in the early innings of AI development. There’s an incredible set of interesting, ongoing challenges to solve: agent knowledge layers, self-improving harnesses, high-quality online and offline evals, and more.

If you want to be at the frontier of AI development and solve hard, industry-defining challenges, join us at Fieldguide, we’re hiring!