cd /news/artificial-intelligence/can-teams-of-ai-agents-recreate-the-… · home › topics › artificial-intelligence › article
[ARTICLE · art-172878] src=news.northeastern.edu ↗ pub= topic=artificial-intelligence verified=true sentiment=↑ positive

Can teams of AI agents recreate the chain reaction of human decisions?

A Nov. 16 study in Nature Computational Science by Northeastern University mechanical and industrial engineering professor Babak Heydari and graduate student Hongan Zhu showed that AI models can reproduce human decision-making more closely by accounting for the way people actually make choices, flaws included. The findings, reported by Northeastern Global News, are aimed at building more realistic agent-based simulations that policymakers and businesses can use to test interventions such as price shifts, evacuation instructions or new laws before applying them in the real world. The study also suggests a pathway for keeping AI agents' collective behavior more human-like and easier for developers to guide.

recommendation

When simulating human systems with AI agents, encode realistic, flawed decision rules rather than idealized rationality so collective outcomes like hoarding or segregation emerge as they do in real populations.

more in playbook: Designing agent systems
by read5 min views6 publishedNov 19, 2026
Can teams of AI agents recreate the chain reaction of human decisions?
Image: News (auto-discovered)

People don’t always play by the book, and AI simulations need to account for human quirks. A Northeastern study shows how doing so can help businesses and policymakers try out ideas before the consequences hit home.

At the start of the COVID-19 pandemic in 2020, people hoarded toilet paper. Australians alone bought 39.7 million rolls from the Woolworths supermarket chain in a single week, according to the store’s 2020 annual report.

It’s not the only time that fear of a shortage spiraled into one.

In 2012, an announcement that Hostess was shutting its doors led to the “Twinkie apocalypse” as people rushed to stockpile their beloved snacks. One Florida man bought 20 boxes of Twinkies and Ding Dongs for about $100 total, then hawked them for $1,000 per box, Wired reported at the time.

“Everyone in the supply chain starts reacting to other parts of the chain,” Northeastern University professor of mechanical and industrial engineering Babak Heydari told Northeastern Global News. This idea applies to customers and suppliers alike.

Accurate simulations of human decision-making, in turn, can help anticipate supply-chain jams and test ways to avoid them.

In a Nov. 16 study in Nature Computational Science, Heydari and graduate student Hongan Zhu showed how AI models could reproduce human decision-making more closely by taking the way people make choices – flaws and all – into consideration.

The findings can help scientists build more realistic simulations of human decision-making across a wide range of situations, Heydari said. Accurately predicting human judgment means research teams can test how people would respond before anyone makes changes in the real world.

Policymakers and businesses can use the simulations to test interventions – changes intended to improve how a system works, such as price shifts or new laws.

For example, officials can test which evacuation instructions help residents reach safety and which ones might funnel drivers into gridlock as they try to leave. Transit planners could test whether lower fares would encourage drivers to take the bus. A business could explore whether moving a pharmacy to a different location would attract new customers or make it harder for existing clients to reach it. Doing so can avoid costly mistakes or strategies that succeed on screen but backfire in practice. The findings also suggest a pathway for keeping AI agents from stepping out of line making their collective behavior more human-like – and easier for developers to understand and guide.

The idea of simulating how people make decisions dates back to a 1971 experiment that hinged on a pocketful of coins and dice. Nobel laureate Thomas Schelling used them to demonstrate how small choices add up to large-scale urban segregation patterns. He placed different coins on a grid to represent residents from two groups. A resident would “move” if fewer than half their neighbors shared their background. Each move changed the balance of groups within neighborhoods and prompted others to relocate. Over time, a mixed community split into separate clusters.

Scientists eventually swapped coins and dice for computers to test these types of scenarios, but the experiment laid the groundwork for agent-based simulation – modeling behavior by proxy, Heydari explained.

“Human studies take time and money, which limits how many experiments we can run in practice,” said Jakob Kaiser, senior researcher at the Nuremberg Institute for Market Decisions, a nonprofit in Germany.

AI offers a shortcut.

Simulations also get around ethical concerns, added Sangeet Khemlani, senior cognitive scientist at the U.S. Naval Research Laboratory in Washington, D.C. AI can play out all sorts of scenarios that would be too risky to test on real people without ending up worse for wear, he said.

That said, simulations have to be accurate in order to be useful. Incorrect predictions can backfire, Khemlani noted.

To test how closely AI could mimic people, Heydari and Zhu got agents from seven AI models to play the so-called beer game, where four agents – a retailer, a wholesaler, a distributor and a factory – run a pretend supply chain in response to customer demand, which is set by the game.

Each player sees only part of the picture and decides how much beer to order or supply. Deliveries take time, so players must also keep track of what’s on the way. Everyone tries to keep enough beer on hand without drowning in surplus. Order too little, and customers go without. Get too much, and storage costs pile up.

Left to their own devices, the agents made some odd choices. Some went overboard even after demand leveled off instead of adjusting as humans might.

Then, Zhu and Heydari added “cognitive grounding” – a process of teaching agents how people estimate demand, respond to shortages or excess and sometimes fail to fully account for supply already on the way. The information came from earlier experiments that analyzed how humans play the beer game.

“It is just a paraphrase of the mathematical model of how to play the game,” Zhu said. “It doesn’t contain anything that explicitly tells them how much they should buy.”

Across 231 sessions, beer-buying bots moved closer to the human benchmark. Some nearly matched it.

“Basically, they internalized what humans would do,” Zhu said.

In another round of testing, the researchers described what people might experience during the game instead of explaining their decisions. For example, players might feel uncertainty, pressure to order more or an urge to overcorrect.

Agents from six of the seven models responded more like people as a result.

That said, some bots were set in their ways. One cut back from ordering 1,000 beers to 100 or 50.

Zhu compared the way their lab approaches AI to psychologists treating patients.

“We don’t go into the neural pathways,” he said. Instead, they observe how the agents behave and use language to redirect the choices they make.

For Heydari, better predictions translate to practical improvements. “Our lab has one leg in engineering and one leg in computational social sciences and economics,” Heydari said. “We care about understanding things, but once we get that understanding, we want to use it to design better ways to improve the systems.”

── more in #artificial-intelligence 4 stories · sorted by recency
── more on @northeastern university 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/can-teams-of-ai-agen…] indexed:0 read:5min 2026-11-19 · —