Anatomy of the LLM-agent wall — what JEPA, ARC-AGI-3 and a $1bn world-model bet say about what comes next

The conversation about agentic AI has split. Anthropic, OpenAI, and most of the agent platforms built on top of them say the path runs through better reasoning, longer context, and smarter scaffolding around a frontier LLM. Yann LeCun, Demis Hassabis, Fei-Fei Li, and most of the world's top early-stage AI capital say the architecture is wrong from the bottom up.

Six months ago this looked like a researcher disagreement. Three things have arrived since then: a benchmark frontier models can't beat, an architectural alternative with theoretical foundations that finally hold up, and a $2bn-plus capital allocation against the current paradigm. Reading them together is the cleanest way to see what 2027's agent stack probably looks like.

ARC-AGI-3 is the cleanest version of a pattern that has shown up in three or four unrelated benchmarks in 2026. Frontier models do not generalise to environments the harness writer hasn't seen, and test-time compute does not fix it.

Symbolica's Arcgentica harness, built on the open-source Agentica SDK, scored 36.08% on the public ARC-AGI-3 set — 113 of 182 levels solved, total bill $1,005. The underlying model is Claude Opus 4.6, the same model that scores 0.2% if you point it at the benchmark directly. The difference is an orchestrator-subagent architecture in which the top-level orchestrator never touches the environment. Subagents do, and they return compressed textual summaries. The orchestrator plans against those summaries, and the context never grows.

Chollet's note in the technical paper: the compression the Symbolica harness performs — keeping a small, useful representation of the world while interacting with it — is "exactly what ARC-AGI-3 is designed to test as a native capability, not an engineered workaround." The current best ARC-AGI-3 result is being produced by a harness doing what the model can't.

General AgentBench, published on arXiv in February 2026, evaluated ten leading LLM agents across a unified search-coding-reasoning-tool-use environment. The authors reported "substantial performance degradation when moving from domain-specific evaluations to this general-agent setting," and went further: "neither scaling methodology yields effective performance improvements in practice, due to two fundamental limitations: context ceiling in sequential scaling and verification gap in parallel scaling." Test-time compute does not climb the wall.

WebArena multi-site, from November 2025, is bleaker. The default success rate of a frontier LLM agent on the standard web-navigation benchmark is 2%. Environment-interaction adaptation lifts it to 23%. The baseline on a benchmark the field has been working on for two years is 2%. The Aegis paper from August 2025 reported the most dispiriting number: OpenAI's o3, optimised specifically for agentic use, improves on customer-service workloads by 2-3% over o1. The marginal return on frontier-model reasoning on the tasks the agent industry sells is approaching zero.

The reason the harness ends up doing the model's job is structural. An LLM agent operates entirely in token space. Every prediction is over a probability distribution of words that describe the world. Every plan is a sampled sequence of words describing actions. Every state is whatever the agent has kept in its context window. Planning is generative text. The world is whatever the most recent few thousand tokens claim it is.

LeCun's argument, in the 2022 position paper A Path Towards Autonomous Machine Intelligence and in half a dozen JEPA papers since: this architecture fights the task. A model that predicts tokens spends most of its capacity on details that are not predictable — lighting, phrasing, the specific path a user takes — and has limited room left for the parts that matter for planning. Predict in a learned latent space rather than observation space. Encode the world into representations. Predict the next representation, conditioned on an action. Choose actions that minimise a learned distance between the predicted representation and a goal representation. No tokens. No language. The planning loop is inside the latent space the model already learned.

For three years it was easy to argue that JEPA was a position paper with no working systems behind it, and that the latent-space architecture was prone to representational collapse — the failure mode where the encoder maps every input to the same constant vector because that trivially minimises any prediction loss. The community kept the framework alive with a pile of training heuristics: stop-gradients, teacher-student networks with EMA targets, predictor heads, whitening layers. Each fix made JEPA harder to scale and easier to dismiss as research-grade rather than production-ready.

The last twelve months closed that gap. Three papers in particular are the ones to cite if anyone says JEPA is still vapour.

The money has moved faster than the consensus. On March 9, 2026, AMI Labs — the startup LeCun co-founded after leaving Meta in November 2025 — closed a $1.03bn seed round at a $3.5bn pre-money valuation. The largest seed round ever raised by a European company. Co-leads: Cathay Innovation, Greycroft, Hiro Capital, HV Capital, Bezos Expeditions. Strategic investors: NVIDIA, Samsung, Toyota Ventures, Bpifrance, Temasek. Individuals: Jeff Bezos, Mark Cuban, Eric Schmidt, Tim Berners-Lee, Xavier Niel.

The team. CEO Alexandre LeBrun, formerly co-founder and CEO of medical-AI company Nabla, where he reached LeCun's conclusion about LLMs from the patient-safety angle. CSO Saining Xie, ex-DeepMind. CRIO Pascale Fung, ex-Meta senior director of AI research. VP World Models Mike Rabbat, ex-Meta FAIR research director. COO Laurent Solly, Meta's former VP for Europe. Hubs in Paris, New York, Montreal, Singapore.

Fei-Fei Li's World Labs raised $1bn in February 2026 and shipped Marble. DeepMind released Genie 3 in August 2025 — the first real-time interactive world model, ~11B parameters, 24fps. NVIDIA's Cosmos foundation models passed 2M downloads. xAI hired ex-NVIDIA specialists for world-model work in late 2025. The aggregate early-stage capital allocated to the "LLMs are not the path" thesis since October 2025 sits over $2.5bn, before counting infrastructure spend.

François Chollet, who runs the ARC Prize Foundation and co-founded Ndea on the parallel program-synthesis bet, endorsed the AMI raise publicly: "AMI Labs proves that you can raise massive capital for fundamental research outside of Silicon Valley." Chollet's program-synthesis thesis and LeCun's JEPA thesis are different architectural bets, but they run the same critique of the LLM-agent stack from opposite ends.

Anthropic's Dario Amodei has been on record since 2024 saying "a country of geniuses in a datacenter as early as 2026" is reachable through LLM-derived systems. The current state of that bet is Claude Mythos — a reported 10-trillion-parameter model that costs roughly 5× Opus 4.6 to run, isn't publicly served, and that Anthropic has framed as a cybersecurity tool. If scaling were producing the returns the 2024 case required, Mythos would be the consumer flagship. Instead it's a partner-API product with no public benchmarks.

If the architectural argument is right, the question of which agent platform to standardise on this year is less about which model leads the benchmarks and more about which platforms survive a paradigm change. None of these tools become useless overnight — LeCun himself has said any new architecture is multi-year — but long-term commitments to specific agent stacks look different if the underlying architecture is provisional.

The world-model camp has unsolved problems of its own. Bessemer's analysis from March 2026 identified the structural one: world models don't batch the way LLMs do. A 70B LLM costs a few cents per hour per user because dozens of users share each chip's throughput. A real-time video-generation world model can't share inference cost across users. NVIDIA's Cosmos training run used 10,000 H100 GPUs over three months. Serving economics may turn out to be the bottleneck even if the architecture is right. The PAN paper (arXiv 2507.05169) made the related point that an LLM backbone may need to stay in the loop alongside a generative latent predictor — the camps may converge on a hybrid stack neither of them currently sells.

LLM agents are not going to disappear. The architecture being marketed today — one big model, a long context window, a tool-calling harness, and the assumption that scaling continues — is the part that probably does. The pieces that survive are the harness expertise, the latent-space planning idea, and the small specialist models that already do narrow work well.