Enterprise teams building multi-agent AI systems may be paying a compute premium for gains that don’t hold up under equal-budget conditions. New Stanford University research finds that single-agent systems match or outperform multi-agent architectures on complex reasoning tasks when both are given the same thinking token budget.

However, multi-agent systems come with the added baggage of computational overhead. Because they typically use longer reasoning traces and multiple interactions, it is often unclear whether their reported gains stem from architectural advantages or simply from consuming more resources.

To isolate the true driver of performance, researchers at Stanford University compared single-agent systems against multi-agent architectures on complex multi-hop reasoning tasks under equal “thinking token” budgets.

Their experiments show that in most cases, single-agent systems match or outperform multi-agent systems when compute is equal. Multi-agent systems gain a competitive edge when a single agent’s context becomes too long or corrupted.

In practice, this means that a single-agent model with an adequate thinking budget can deliver more efficient, reliable, and cost-effective multi-hop reasoning. Engineering teams should reserve multi-agent systems for scenarios where single agents hit a performance ceiling.

Understanding the single versus multi-agent divide

Multi-agent frameworks, such as planner agents, role-playing systems, or debate swarms, break down a problem by having multiple models operate on partial contexts. These components communicate with each other by passing their answers around.

While multi-agent solutions show strong empirical performance, comparing them to single-agent baselines is often an imprecise measurement. Comparisons are heavily confounded by differences in test-time computation. Multi-agent setups require multiple agent interactions and generate longer reasoning traces, meaning they consume significantly more tokens.

Consequently, when a multi-agent system reports higher accuracy, it is difficult to determine if the gains stem from better architecture design or from spending extra compute.

Recent studies show that when the compute budget is fixed, elaborate multi-agent strategies frequently underperform compared to strong single-agent baselines. However, they are mostly very broad comparisons that don’t account for nuances such as different multi-agent architectures or the difference between prompt and reasoning tokens.

“A central point of our paper is that many comparisons between single-agent systems (SAS) and multi-agent systems (MAS) are not apples-to-apples,” paper authors Dat Tran and Douwe Kiela told VentureBeat. “MAS often get more effective test-time computation through extra calls, longer traces, or more coordination steps.”

Revisiting the multi-agent challenge under strict budgets

To create a fair comparison, the Stanford researchers set a strict “thinking token” budget. This metric controls the total number of tokens used exclusively for intermediate reasoning, excluding the initial prompt and the final output.

The study evaluated single- and multi-agent systems on multi-hop reasoning tasks, meaning questions that require connecting multiple pieces of disparate information to reach an answer.

During their experiments, the researchers noticed that single-agent setups sometimes stop their internal reasoning prematurely, leaving available compute budget unspent. To counter this, they introduced a technique called SAS-L (single-agent system with longer thinking).

Rather than jumping to multi-agent orchestration when a model gives up early, the researchers suggest a simple prompt-and-budgeting change.

“The engineering idea is simple,” Tran and Kiela said. “First, restructure the single-agent prompt so the model is explicitly encouraged to spend its available reasoning budget on pre-answer analysis.”

By instructing the model to explicitly identify ambiguities, list candidate interpretations, and test alternatives before committing to a final answer, developers can recover the benefits of collaboration inside a single-agent setup. 

The results of their experiments confirm that a single agent is the strongest default architecture for multi-hop reasoning tasks. It produces the highest accuracy answers while consuming fewer reasoning tokens. When paired with specific models like Google’s Gemini 2.5, the longer-thinking variant produces even better aggregate performance.

The researchers rely on a concept called “Data Processing Inequality” to explain why a single agent outperforms a swarm. Multi-agent frameworks introduce inherent communication bottlenecks. Every time information is summarized and handed off between different agents, there is a risk of data loss.

In contrast, a single agent reasoning within one continuous context avoids this fragmentation. It retains access to the richest available representation of the task and is thus more information-efficient under a fixed budget.

The authors also note that enterprises often overlook the secondary costs of multi-agent systems.

“What enterprises often underestimate is that orchestration is not free,” they said. “Every additional agent introduces communication overhead, more intermediate text, more opportunities for lossy summarization, and more places for errors to compound.”

On the other hand, they discovered that multi-agent orchestration is superior when a single agent’s environment gets messy. If an enterprise application must handle highly degraded contexts, such as noisy data, long inputs filled with distractors, or corrupted information, a single agent struggles. In these scenarios, the structured filtering, decomposition, and verification of a multi-agent system can recover relevant information more reliably.

The study also warns about hidden evaluation traps that falsely inflate multi-agent performance. Relying purely on API-reported token counts heavily distorts how much computation an architecture is actually spending. The researchers found these accounting artifacts when testing models like Gemini 2.5, proving this is an active issue for enterprise applications today.

“For API models, the situation is trickier because budget accounting can be opaque,” the authors said. To evaluate architectures reliably, they advise developers to “log everything, measure the visible reasoning traces where available, use provider-reported reasoning-token counts when exposed, and treat those numbers cautiously.”

What it means for developers

If a single-agent system matches the performance of multiple agents under equal reasoning budgets, it wins on total cost of ownership by offering fewer model calls, lower latency, and simpler debugging. Tran and Kiela warn that without this baseline, “some enterprises may be paying a large ‘swarm tax’ for architectures whose apparent advantage is really coming from spending more computation rather than reasoning more effectively.”

Another way to look at the decision boundary is not how complex the overall task is, but rather where the exact bottleneck lies.

“If it is mainly reasoning depth, SAS is often enough. If it is context fragmentation or degradation, MAS becomes more defensible,” Tran said.

Engineering teams should stay with a single agent when a task can be handled within one coherent context window. Multi-agent systems become necessary when an application handles highly degraded contexts. 

Looking ahead, multi-agent frameworks will not disappear, but their role will evolve as frontier models improve their internal reasoning capabilities.

“The main takeaway from our paper is that multi-agent structure should be treated as a targeted engineering choice for specific bottlenecks, not as a default assumption that more agents automatically means better intelligence,” Tran said.

from the making-America-late-on-interest-payments-again dept

The government needs more funding than ever, which is kind of hilarious when you realize the Tea Party of the Obama era was the predecessor of this Big Government version of the GOP.

The DHS can’t even get itself a budget at the moment. Sure, it will get some money thrown to it sooner or later and the administration won’t let the lack of tax revenue offsets stop it from feeding billions more into its Bigotry Machine.

But that’s not all. Behold our all-but-officially-declared war in Iran, currently headed by the Department of Defense War Little Excursion, which is adding billions of dollars weekly to the national deficit. After all, as right-leaning libertarians like to point out, the government doesn’t actually “make” anything. The private sector builds the bombs and missiles. And unlike TSA agents, they expect to be paid.

You know who could help this country offset some of its insane expenditures? It’s the same people we’re spending billions to remove from the country:

Immigrants accounted for more US income and generated more revenue for the government because they were, on average, over 12 percentage points more likely to be employed than the US-born population. This means that even if immigrants earn lower hourly wages, they can still account for more total income per capita than the US-born population by working cumulatively more hours. This higher employment rate was driven by the fact that immigrants were, on average, 20 percentage points more likely to be of working age. Immigrants usually arrive in the US as young adults and often leave before retirement.

More succinctly, immigrants out-punch their weight class when it comes to erasing budget deficits:

Accounting for savings on interest payments on the national debt, immigrants saved $14.5 trillion in debt over this 30-year period.

[…]

Without the contributions of immigrants, public debt at all levels would already be above 200 percent of US GDP—nearly twice the 2023 level and a threshold some analysts believe would trigger a debt crisis.

But that help is apparently no longer welcome. The Trump administration has succeeded in eliminating the firewall between the IRS and ICE, allowing ICE agents to use this data to hunt down taxpayers who work harder and pay more taxes than the white, natural-born citizens that this administration pretends make America great.

That’s going to cause even more problems for an administration that is spending far more liberally than any “liberal” it blames its current budget problems on. Here’s how that looks on the ground as Tax Day has come and gone in the United States:

By the time Tax Day rolls around every April 15, accountant María José Solís usually has more to do. More clients. More paperwork. More phones ringing, more emails and WhatsApp messages pinging.

But this year, she said, more than 550 of her regular clients have disappeared. That’s about 15 percent of her customer base at Toro Taxes, the bilingual firm in Wheaton, Maryland, that Solís runs.

There’s your anecdote, albeit one that’s being repeated around the nation. Here’s the data:

The Yale Budget Lab estimates that the IRS stands to lose between $147 billion and $479 billion over the next decade as migration to the U.S. declines, deportations increase and immigrants of various statuses disengage from the formal economy for what some experts say may be an extended period.

That estimate will likely be low if the Trump administration continues to purge migrants at the rate it has since Trump returned to office. It will definitely be lower if another similarly bigoted GOP lawmaker succeeds him as president.

And it’s not just the losses up front. There’s money leaking out the back as well. It’s a double-dip, because migrants with ITINs (individual tax identification numbers) pay taxes for services they can’t actually access, like Social Security and Medicare. They’re actually subsidizing citizens who pay fewer taxes, work fewer hours, and commit more crimes than they do.

This nation continues to become poorer, not just in terms of financial viability, but in heart and spirit. Migrants made this nation great. Now, a bunch of ungrateful people who hate people who aren’t white are not only driving us deeper into debt, but they’re eliminating a source of income that never asked for anything more than a chance to survive.

Filed Under: bigotry, cbp, dhs, ice, immigration, irs, mass deportation, trump administration

from the good-deals-on-cool-stuff dept

Dive into Godot – a rising star in the game engine world – with the 2026 Complete Godot Stack Development Bundle. You’ll learn to create platformers, RPGs, strategy games, FPS games, and more as you master this free and open-source engine with easily expandable systems. Plus, you’ll also explore techniques for game design and game asset creation – giving you the ultimate techniques to customize your projects. It’s on sale for $25.

Note: The Techdirt Deals Store is powered and curated by StackCommerce. A portion of all sales from Techdirt Deals helps support Techdirt. The products featured do not reflect endorsements by our editorial team.

Filed Under: daily deal

from the but-maybe-this-time-it-will! dept

Back in 2011 and 2012, one of the central technical objections that helped kill SOPA and PIPA was about DNS blocking. Engineers, internet architects, and cybersecurity experts all lined up to explain, in painstaking detail, why blocking at the DNS layer was a terrible idea. It would break the fundamental architecture of how the internet works. It would have massive collateral damage. It would undermine security protocols designed to protect users from exactly the kind of DNS manipulation that the bill proposed. And it wouldn’t even stop piracy, because anyone who actually wanted to get around DNS blocking could do so easily.

Congress, to its rare credit, actually listened to the technical experts (and widespread protests) and shelved the legislation. But the entertainment industry never gave up on the idea. They just went jurisdiction-shopping. And France, which has never met a maximalist copyright enforcement scheme it didn’t love, has been more than happy to oblige.

As recently reported by TorrentFreak, a Paris Court of Appeal validated DNS blocking orders requiring Google, Cloudflare, and Cisco to block access to pirate sites through their own DNS resolvers. This goes beyond traditional ISP resolvers, which France has been ordering blocked for years — this targets third-party resolvers — the ones that millions of people specifically choose to use because they offer better privacy, better security, and better reliability than their ISP’s default DNS.

But, of course, in France (and to the usual crew of Hollywood lobbyists), “better privacy, security, and reliability” can only mean one thing: used for piracy.

The court rejected all five appeals, and in doing so, articulated a legal principle so sweeping that it has no natural stopping point.

In this case, French pay-TV provider Canal+ went to court under Article L. 333-10 of the “French Sport Code,” which lets rightsholders request “all proportionate measures” against “any online entity in a position to help” block access to pirate sites. Canal+ argued that because users were simply switching to third-party DNS resolvers to circumvent ISP-level blocking, those resolvers should be conscripted into the blocking regime too.

Cloudflare and Cisco pushed back, arguing that their DNS resolvers serve a “neutral and passive function” — they translate domain names into IP addresses and that’s it. They compared their role to a phone book. The court’s response boiled down to: we don’t care.

The DNS resolution service allows its users, via the translation of a domain name into an IP address, to access websites on which sports competitions are broadcast in violation of rights-holders’ rights, and in particular to circumvent the blocking of those sites by ISPs.

The court found that the “neutral and passive” nature of DNS resolvers is “simply irrelevant to Article L. 333-10.” The law isn’t about liability at all — it only cares whether a service can help block access to pirate sites, which DNS resolvers clearly can. If you are technically capable of blocking access, you must.

Google, meanwhile, tried a different argument: that DNS blocking through third-party resolvers isn’t effective because users can just switch to a VPN or yet another resolver. The court wasn’t moved by that either:

Any filtering measure can be circumvented, and this possibility does not render the measures in question ineffective.

As long as DNS blocking stops some subset of users from reaching pirate sites, the court ruled, it’s “proportionate.” Under that line of thinking, any measure that inconveniences even a fraction of would-be pirates is legally justified, no matter how much collateral damage it causes for everyone else.

And if you think that principle has any limit, Canal+ has made it quite clear that they don’t think it does:

Canal+ said in a statement that the rulings are “more than a victory,” forming part of “a global approach that will be reinforced by the progressive deployment of complementary measures, including IP blocking.”

Canal+ has already been getting courts to order VPN providers to block as well. So now we have ISP DNS blocking mandated, third-party DNS resolver blocking mandated, VPN blocking mandated — and, per the TorrentFreak article, direct automated IP address blocking is coming too. They will not stop until the entire internet is broken.

Each step reaches further down the internet stack, breaks more of the internet for more people, and stops fewer actual pirates, because the people who are determined to pirate content are always one technical maneuver ahead. The people who get caught in the collateral damage are ordinary users who happen to use Cloudflare’s 1.1.1.1 or Google’s 8.8.8.8 for perfectly legitimate reasons like speed, reliability, and privacy.

Cisco, rather than comply with the original order, simply pulled its OpenDNS service out of France entirely. That’s the kind of collateral damage we’re talking about. French users who relied on OpenDNS for entirely lawful purposes completely lost access to the service. Because a copyright holder decided that the DNS layer was the right place to play whack-a-mole with pirate sites.

When Cisco argued on appeal that implementing geo-targeted DNS blocking would require 64 person-weeks of engineering work, the court waved it off, saying the estimate was “not supported by any objective evidence” and pointing out that Cisco already offers DNS filtering to enterprise customers. The fact that enterprise DNS filtering for corporate networks is a fundamentally different thing than mass geo-targeted blocking of domains at the resolver level for an entire country’s users apparently did not register as a meaningful distinction.

The court’s core reasoning — that any entity technically capable of blocking must do so, that circumvention doesn’t make blocking disproportionate, and that the “neutral and passive” function of an intermediary is irrelevant — creates a legal framework that can reach basically anything. If a DNS resolver can be conscripted because it’s “in a position to help,” what about browsers? What about operating systems? What about CDNs, or cloud hosting providers, or certificate authorities? The logic has no brake pedal. Every layer of the internet stack is, in some sense, “in a position to help” block access to content. The question the court’s reasoning cannot answer is: where does it end?

Under this reasoning, what’s to stop a rightsholder from arguing that browsers should block pirate URLs directly? Or that operating systems should refuse to resolve them at all?

That seems bad!

Of course, this kind of maximalist copyright enforcement is something of a French specialty. This is the same country that brought us HADOPI, the graduated response agency that cost French taxpayers €82 million over a decade while imposing a grand total of roughly €87,000 in fines. A staggering return on investment — if the goal was to light money on fire while accomplishing nothing. France has also been at the forefront of copyright exceptionalism that risks undermining the EU legal system more broadly, pushing interpretations of copyright law so aggressive that they threaten to distort the legal frameworks of neighboring countries.

France keeps doing the same thing over and over again: spend enormous sums, conscript more and more intermediaries, break more and more of the internet’s infrastructure, accomplish almost nothing in terms of actually reducing piracy, and then conclude that what’s really needed is… more of the same, but harder. The entertainment industry’s refusal to learn from twenty years of evidence that enforcement-maximalism doesn’t work is genuinely remarkable. Every study and every natural experiment shows the same thing: the most effective anti-piracy tool ever invented is convenient, reasonably priced legal access to content. But that requires adapting your business model, and it’s apparently much more satisfying to get courts to break the internet for you instead.

The ruling’s real danger is the template it sets. Other countries with similar legal frameworks will look at this appeals court validation and think: we can do that too. The “any entity in a position to help” standard, combined with the “doesn’t have to be perfectly effective” standard, combined with the “we don’t care about your neutral role in the architecture” standard, adds up to a legal toolkit for conscripting nearly any internet infrastructure provider into a copyright enforcement apparatus. And the costs get externalized onto those providers (and their users), while the rightsholders collect the benefits.

The engineers who fought SOPA warned about exactly this: DNS blocking breaks things, creates collateral damage, pushes enforcement into layers of the stack never designed for it — and doesn’t actually stop piracy, because the actual pirates just route around it while everyone else suffers. France apparently decided all of those concerns are, to quote the court, “simply irrelevant.” And now they’ve moving on to IP blocking.

At some point, you run out of layers of the internet to break. But apparently we’re going to have to find out where that point is the hard way.

Filed Under: copyright, dns blocking, france, sopa

Companies: canal plus, cisco, cloudflare, google

• Assassin’s Creed Black Flag Resynced will feature PSSR 2 on PS5 Pro
• Ubisoft confirms the game will have several modes on PS5 and PS5 Pro
• The game will also use the latest version of the studio’s Anvil engine

Ubisoft has confirmed Assassin’s Creed Black Flag Resynced will feature Sony‘s upgraded PSSR (PlayStation Spectral Super Resolution) AI upscaling technology.

The remake of the 2013 Assassin’s Creed pirate adventure has finally been revealed, and a new PlayStation Blog post following the world premiere trailer has outlined what players can expect from the PS5 and PS5 Pro versions.

Back in the early 1980s, Radio Shack sold a variety of gadgets that caught the eye of both curious children and tech enthusiasts. Among those goods was Gobble Man, a small handheld game that placed maze chases directly into your palms years before the larger portable consoles that came later. Bandai first released this in Japan during 1981 as a game known as Packri Monster. Tandy then scooped it up, licensing the design for its US stores before selling it as Gobble Man in 1983. To add to the confusion, Tandy sold the exact same units under the titles Hungry Monster and Ogre Eater.

Four AA batteries powered everything from start to finish. A vacuum fluorescent display, which was basically a tiny little screen, made the game feel extremely bright; even in a dark room or on a vehicle journey, you could see every wall and dot in the maze clearly. Inside the handheld was a Hitachi HD38800 chip, a specialized microcontroller designed specifically to operate this game and nothing else.

PlayStation 5 Portal Remote Player – Midnight Black

• Play Your Game Collection with Remote Play – PlayStation Portal Remote Player can play compatible games you have installed on your PS5 console…
• Cloud Streaming from the Game Catalog and Classics Catalog – Discover an awesome library of PS5 games on the PlayStation Portal Remote Player with…
• Cloud Streaming for PS5 Games in Your Library – With PlayStation Plus Premium, stream select digital PS5 games in Your Library from PlayStation Store…

Players took a rectangular plastic case about the size of a large paperback book to play with. A four-way joystick on the front let you to direct the yellow monster through the curving courses; the stick itself employed a dual pivot mechanism that required a forceful push to register every turn. The monster continued to move in the same direction until the player altered course or it collided with a wall. The controls included a start button and a power switch, as well as a small orange indicator that signaled when the gadget was turned on.

The music came from a very small built-in speaker that played the tunes and effects at the same constant volume, such as beeps and rudimentary melodies, and alerted you when something happened in the game. There was no volume control, so the cacophony was part of the experience, whether you liked it or not. The gameplay took place on a single fixed maze, with green food pellets to chew on. The goal was simple: move the monster so that it ate every pellet before the bogey caught up, and if it did, game over. Red power food would sometimes materialize in specific locations. Eating one of those would transform the monster into a coward who would back direction and flash briefly. For a certain period of time, you might pursue them and get some bonus points.

However, the difficulty level increased with each subsequent round. More enemies joined the chase, and speeds rose, giving the maze the sensation of closing in around you when the strain was on. Your score would skyrocket if you timed your power snacks perfectly and wiped the board clear without making a single error. Round after round, the whole thing just kept going until the batteries died or your thumbs gave out in frustration. Anyone who picked up Gobble Man in 1983 would immediately feel as if they understood the game because the rules were so familiar.

Gobble Man was never really trying to compete with the big boys, such as Pac-Man, home consoles, and coin-op machines. It was simply a nice little distraction to keep you engaged while on the run, regardless of where you were. Six years later, Nintendo released the Game Boy, which very well defined what portable gaming was all about. But first, Tandy created Gobble Man, which proved that the concept worked and left many people wanting more.

More than half of the outlets listed on its website are reportedly closed

Homegrown supermarket and minimart chain Hao Mart has closed a large share of its outlets as losses deepen, The Straits Times reported.

As of publication, Hao Mart’s website lists 20 outlets across Singapore. However, a check by The Straits Times found that only seven remain in operation after visiting all listed locations over a two-week period in the second half of Mar.

The remaining seven include six regular Hao Mart stores, located in Bedok South Avenue 3, Canberra Link, Potong Pasir Ave 1, Petir Road, Whampoa Drive and Pasir Ris Street 21; and one premium Eccellente outlet in the Marina Square shopping mall.

According to Accounting and Corporate Regulatory Authority (ACRA) records obtained by the publication, the company has been in the red since 2023, and its losses have worsened over the years.

For FY2025, Hao Mart reported a S$49.6 million loss. The amount is up from S$32.8 million in FY2024 and S$23.2 million in FY2023, following two years of profitability. The company also posted a S$2.2 million loss for FY2019, its first year of filings.

FY2025’s figures were filed with ACRA belatedly in Jan. ACRA told The Straits Times in Mar that it has taken enforcement action against Hao Mart for failing to file annual returns within six months of its financial year-end.

While the regulator did not disclose the specific penalty imposed, its website states that late filings can incur fines of S$300 or S$600, depending on the length of the delay.

Separately, Hao Mart is also facing four High Court lawsuits, including a dispute with landlord OG over the termination of its lease for its flagship five-storey Taste Orchard mall on Orchard Road.

Vulcan Post has reached out to Hao Mart about the scaling down of its operations and financial situation.

Some closed stores taken over by rivals

Hao Mart was founded by Dr Tan Kim Yong in 2016, with its first store at Block 74 Whampoa Drive. The chain expanded rapidly over the years, peaking at 51 stores in Dec 2021.

But by Dec 2024, Hao Mart stores had dwindled to just 20, according to its website. Among the outlets that The Straits Times found to be closed, four had been replaced by Hao Mart’s rivals.

Sheng Siong now occupies Hao Mart’s former Punggol East and KINEX Mall spaces, while ValueMart runs the Punggol Walk site near Waterway Point.

Newcomer ACE Signature took one of three Geylang shophouses. Of the remaining two Geylang units, one became a coffeeshop, and the other was demolished in Sep 2024 for a five-storey rebuild. The Redhill Road minimart is now a tuition centre.

Elsewhere, Indian supermarket chain Sri Murugan Supermarket now runs Hao Mart’s former Bayshore Park condo minimart. Its Parksuites condo-facing outlet has been converted into an after-school care centre. The Far East Plaza outlet is now a food court, the Esplanade Xchange space has become a travel agency, and the East Village supermarket has been replaced by a gym.

• Read other articles we’ve written on Singaporean businesses here.