Gabriel Denis-Arrue Munes is a Product Manager at OpenSpace, building the agent sandbox
In our recent blog posts, we covered what AI agents are and why construction, more than any document-heavy industry, needs agents that can see. This post gets specific. It’s about site search agents: what they do, how they work, and what becomes possible when they run in the background while your team is doing everything else.
Site search agents getting to work
A site search agent takes a question—check Level 3 for missing PPE, find the toilets on Level 10, flag any rooms where ductwork hasn’t started yet—and answers it against the visual record of a project. The agent can run on demand when a PM asks, when a new capture finishes processing, on a schedule overnight, or trigger across an entire portfolio at the start of the week.
Every member of a project team can have several agents working on their behalf around the clock. The visual record from yesterday afternoon’s walk, for example, will have already been examined by the time anyone is at the trailer the next morning. That changes the math on what a team can get done in a day.
Think of this agent as a tireless PE. One you can send to run through your job, at any time, as often as you want.
What async agents like site search can look like
It’s 5 am. The Level 3 capture from yesterday afternoon (which no one had time to look at before heading home) was processed in about 15 minutes, as usual. This means it was ready for async site search agents to do their thing, so that by the time the trailer opens:
- The safety agent scanned every new pano on Level 3 for uncapped rebar and exposed fall hazards. It identified these issues because they’re ones you asked it to look for. It’s not flagging the 300 other issues an LLM would generally identify if you prompted it to “identify every safety issue on this jobsite.”
- The daily report agent packaged all your superintendent’s notes for the day into a clean site diary, ready for your review (likely done even before you got home last night).
- The housekeeping agent filed three issues with the cleaning vendor for blocked egress paths.
- The owner update agent compiled progress photos from this week’s activity into a deck for tomorrow’s owner meeting.
By the time the team gets to standup, the first hour can go to decisions. The legwork is already done. And that’s just one morning, on one floor, with four agents running in the background.
The opportunity to get ahead
Better search—finding the photo you lost, surfacing the Field Note you forgot, locating the spot on a sheet from last Tuesday—saves real time on real problems. But the value has a ceiling. Teams have been managing without perfect search for the history of construction, and each new improvement helps slightly less than the last. The hour you save on today’s search is an hour saved; it doesn’t compound into something larger.
Async agents are different. Every hour your team is asleep, in a meeting, on PTO, or buried in another scope, your agents could be running safety scans, progress checks, punch list closures, or owner-meeting prep. A superintendent has eight to twelve productive hours a day. An agent has 24. Three agents working on a super’s behalf is the operating throughput of a small team. Ten people on a project, three agents each, and your effective delivery capacity stops being a function of headcount.
Day to day, the gap shows up as a few safety findings caught earlier and a punch list that moves to closeout a week sooner. Over the life of a project, it shapes the entire operating tempo. Across similar scopes in a portfolio, it determines which teams set the bar that everyone else has to match.
The harness is the hard part
The intelligent system around the vision model that turns a customer’s question into a reliable answer is the hard part. It’s what we call the agent “harness.” You can think of it like scaffolding. And think of the model as a very smart specialist who can look at one photo at a time. The scaffolding is everything around that specialist: deciding which photos to show out of hundreds of thousands, making sure those photos are good enough to reason on, bringing in information from other resources such as documents and models, referring to past “meta-reasoning” around images, location, and time, stitching the answers together, and knowing when to say “I’m not sure.” Without that, you have a demo. With it, you have something a superintendent will actually let run.
Asynchronous work becomes possible once the system is working. Without it, agents running unsupervised in the background are worse than useless—they’re a liability.
Getting an agent to that bar is genuinely hard, and most of the work is invisible from the outside. I won’t delve into the specifics here, but in short, it takes a lot of the right visual data with accurate location context—applied carefully across sampling, ranking, reasoning, and evaluation—before an agent earns the right to run unsupervised on a customer’s project.
Trust here gets built one run at a time. PMs and PEs who watch an agent get it right for a couple of weeks straight will leave it running. Three false positives in the first morning, and they won’t. A synchronous tool just needs to be useful. An asynchronous one needs to be useful enough that nobody has to check it. That’s a different bar, and it’s the one OpenSpace is climbing.
Where things are today
A site search agent runs in our development environment. It accepts natural-language queries, finds relevant images to answer the question, runs the search, and returns positive hits with bounding boxes and on-sheet locations. The Field Notes review agent—site search’s closest sibling—has begun rolling out to a small group of early-access customers, and the lessons from that work are shaping how site search develops.
The rough edges that matter are the ones affecting accuracy and precision. Across a growing list of customer use cases—safety scans, progress checks, object searches, owner-meeting prep—the agent needs to be reliable enough that PMs, PEs, and superintendents will leave it running without supervision. That’s what the next several months are about.
What’s next
Site search is the first agent customers asked us for, and the one we’re building right now. The Field Notes review agent is already with early-access testers. And more agents are on deck. What we’re targeting across all of them is the same: high-performance, reliable, asynchronous vision agents that understand each customer’s context.
A year from now, some project teams will be running dozens of agents on every capture. We’d like yours to be one of them. If you’re interested in seeing our work in action or joining an early access group, reach out to your OpenSpace representative.
