😺 Watch: AI can do your taxes now

2026-07-01 · Source: The Neuron · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Robotics & Autonomous Systems · Depth: Intermediate, extended

Summary

OpenAI, in collaboration with Thrive Holdings, developed Tax AI, a Codex-powered agent designed to streamline complex tax preparation workflows. Released on July 01, 2026, this system processes diverse data formats including PDFs, spreadsheets, and client notes to prepare tax returns for accountant review. A key innovation is its self-improvement loop: expert practitioners correct the agent's output, transforming these corrections into structured signals. Codex then investigates relevant traces and evaluations, enabling engineers to implement targeted fixes. This iterative process has resulted in processing 7,000 returns, saving a third of preparation time, and achieving approximately 97% draft accuracy, demonstrating how AI agents can integrate effectively into expert-driven tasks by learning from human feedback.

Key takeaway

For AI Engineers developing agents for expert workflows, prioritize building robust feedback loops over solely optimizing model performance. Your system must capture expert corrections as structured signals, enabling measurable product improvements and fostering trust. This "harness engineering" approach ensures agents get safer and more useful, transforming human review into continuous learning rather than lost effort, which is critical for high-stakes applications like tax preparation.

Key insights

AI agent reliability stems from structured expert feedback loops that transform corrections into measurable product improvements.

Principles

• Harness engineering drives agent improvement.
• Agents must provide source evidence.
• Expert corrections yield measurable signals.

Method

The method involves starting with bounded tasks, preserving evidence, capturing expert corrections as patterns, converting these into measurable evaluations, and integrating engineer review for scoped fixes.

In practice

• Prioritize building evaluation metrics.
• Capture user corrections as structured data.
• Implement source citation for AI outputs.

Topics

• OpenAI
• AI Agents
• Tax Preparation
• Codex Model
• Feedback Loops
• Harness Engineering
• Workflow Automation

Best for: Executive, AI Architect, Machine Learning Engineer, AI Engineer, MLOps Engineer, Director of AI/ML

Related on AIssential

• Faster Code, Slower Understanding: A Survival Note from a Year of Agent-First Engineering — AI shifts the engineering bottleneck from code production to comprehension and maintenance, demanding new optimization strategies.
• What is an AI agent harness? — AI agent harnesses are critical for enabling LLMs to execute complex tasks by connecting reasoning to action safely and reliably.
• Production-grade AI agents for financial compliance: Lessons from Stripe — Production-grade AI agents can significantly boost compliance efficiency while preserving human control and auditability.
• Harvey Drives Legal Agent Learning Via ‘Harness Engineering’ — Harness engineering and autoresearch significantly boost legal agent performance through iterative self-improvement and environmental feedback.
• Building Reliable Agentic AI Systems — Agentic RAG systems require robust context and harness engineering for reliability and trust in regulated environments.
• What's Easy Now? What's Hard Now? — AI coding agents' long-term capabilities are primarily determined by the availability and effectiveness of feedback loops, not just open-loop model behavior.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Neuron.