Why I built it
The Atira EHL hackathon challenge was to turn a public tender into one signed bid, working across four websites in 20 minutes with no human in the loop. That’s exactly the kind of manual office work I think agents should take over.
What I built
Finn, Yazan and I built AI agents that read the tender, find a fitting material in the ERP for every requirement, work out the price and sign the bid with a TAN they pull out of the mailbox. Most of the core was mine, including the price calculation, the submission, the main agent and most of the tests. Yazan built the supplier portal agent and Finn made everything cheap and fast. A script tested every branch every five minutes, and a change only went in if the score didn’t drop.
The hard part
Our worst bug made the system skip a tender it should have bid on. A requirement changed while a second search was still double-checking a no-match, the search kept looking for the old values, and we scored 5 out of 100. We fixed it by handing the new values to the search that was already running. Models also like to fill gaps. One agent sent back “NOT EXTRACTED” as the TAN, so now agents may only use facts from pages they saved word for word, and the math never goes through the model.
What I learned
Models are surprisingly good at finding their way around a website they’ve never seen. They’re bad at noticing when they didn’t find something. That’s why most of our work went into checks.
Results
We won 1st place out of 32 teams. On websites the system had never seen, it scored 100 out of 100, at about 8 cents and two minutes per bid.