Looping Agents and AI Workflows

Matt Wolfe of Bidlo Lite explains the current conversation around looping AI agents and why it matters for teams watching AI/ML trends. The bid bond form is used as a practical construction example, not as the whole point of the video.

From there, the discussion widens into agent loops, human bottlenecks, model access, regulation, open-source models, and how these trends may show up in heavy civil and public works bidding.

• 00:00 Introduction to Looping Agents in Construction AI

• 01:30 Manual to Automated Bid Bond Form Completion

• 04:00 Agent Review and Removing Human Bottlenecks

• 06:00 Scaling Automation Across Multiple Projects

• 08:00 Removing Humans from the Loop Entirely

• 10:00 Anthropic’s AI Models and Industry Trends

• 13:30 Government Regulation and AI Model Access

• 15:00 China’s Advancements in Open Source AI Models

• 17:30 Future of AI in Construction and Final Thoughts

Hey, everyone. I am Matt, one of the developers at BIDLO. We're going to start doing a weekly or maybe twice-monthly explainer about what we're seeing in AI and machine learning technology, focusing on big concepts and how BIDLO is thinking about applying these to heavy civil public works construction. One topic I want to start with is "looping," which has been top of mind recently and has generated a lot of activity on Twitter. Anthropic published a really interesting article about this concept, and we fully believe this is the direction these technologies are heading. Many users haven’t yet thought about using agents or large language models (LLMs) in this way.

Anthropic put together a great article, and I want to walk through it in the context of heavy civil construction. We'll use a simple example—a bid bond form—which is generic and easy to understand. We'll also discuss how we're approaching building organizations around this looping architecture. In the past, if you were filling out a bid bond form, you, as a person, would answer questions like: "What’s the liquidated damages on this project?" You would ask an agent where to find that information or have it pull details from the plans. Then you’d enter the information into the form. The next step is to remove yourself from that loop. Instead of filling out the form, you ask the agent to handle it.

The agent goes through the form, extracts each answer from the plans, fills out the bid bond form, and then presents you with the completed form. This automates more of the process. Then, what do you do next? You check whether the information the agent provided—like liquidated damages or working days—is correct by reviewing it against the plans. Naturally, you ask why you have to do that. Why not have another agent review the first agent’s work? Now, once the first agent completes the form, it triggers a second agent to review it. You, as the human, kick off the process, and when it’s done, you review their responses. That’s great but there’s still a human in the loop.

The trend, and what we are aiming for is to remove the human from this process altogether. Why have a human involved in filling out or reviewing the bid bond form at all? The next step is scaling this. Instead of doing this for one project at a time, at the end of the week, you have a list of projects on your bid schedule. You tell an agent to fill out all the bid bond forms for every project on your list, review them, and save them to their respective project files. This shifts the work from a sequential, human-driven process to a parallel, agent-driven one. If you have, say, five projects, five agents work simultaneously to fill out, review, and file the bid bond forms. That is where we are today.

Many developers and legal experts are still at this stage, where a human is part of the loop. As long as a human is involved, they are the bottleneck. Therefore, there’s a big push to remove the human whenever possible, creating a loop that runs autonomously without human blockers. For example, Anthropic illustrated a system where you never have to think about a bid bond form again. You only manage your bid list. Every few minutes, an agent checks your bid list, sees if a bid bond form is missing, fills it out, reviews it, and files it automatically. You can apply this to many tasks beyond bid bond forms—submittals, quotes, and other busy work triggered at the moment you decide to bid on a project.

The agents handle the work and reviews autonomously. When you log in the next day, the agents report all the completed work. You then review the information to decide if it’s good or bad. If it’s bad, you adjust the loop. Anthropic’s article focuses on recursive learning—agents that teach themselves—and removing humans from the workflow. They acknowledge this approach has risks and have called for more government intervention. We are fully convinced this is how these tools should be integrated into workflows in construction. We’re actively working with our customers to think this way.

On the broader ecosystem, Anthropic’s most capable model, Fable, was briefly released to the public in a limited, restricted form but then pulled back due to concerns from the U.S. Government. There were claims, including from Amazon engineers, that safeguards in the model could be bypassed. Whether true or not, it reflects the political and marketing dynamics at play in the U.S. AI space. Fable is a very good model, slightly better than GPT 5.5 in some tests. But these political issues are partly due to competition and the players involved, including Amazon—which, while not an LLM provider, is an infrastructure player competing in the space. We believe many of these debates are marketing-driven.

The models are powerful tools, but they won’t suddenly destroy the world. Another big point is China is catching up to if not surpassing, the U.S. In model capabilities. Currently, the leaders are Anthropic and OpenAI, with products like GPT. But China has made incredible progress with far less computing power. Their models are more efficient and open source, meaning anyone with proper hardware can run them. Recently, ZAI released GLM 5.2, an open-source model with capabilities near the lower end of top GPT and Anthropic models. This democratization is where the industry is inevitably headed. Like in software, the best technologies eventually become open source, allowing anyone to host and run powerful models.

This can be scary when anyone might have unrestricted access to models with Mythos-level capabilities and no safeguards. But it’s inevitable. China’s progress is impressive, especially given compute limitations, but it is also a natural progression as U.S. Models plateau in capability. The bottleneck is no longer the models themselves but the harnesses and environments that make them effective in practical applications. Because of this plateau, Chinese models are catching up, which contributes to the market noise around issues like Fable being pulled. My point is not to suggest anyone is acting maliciously but to encourage critical reading of headlines—they are often hype.

The difference now is that these models are approaching their current ceiling, and ongoing improvements are driven more by better tooling and environments rather than raw model capabilities. We continue to monitor these developments daily to understand how best to apply this technology at BIDLO and for our customers. If you have any questions or want to learn more, let me know. Otherwise, I’ll see you in the next explainer.