What changed: from bounded tasks to multi-year competitions
The old picture: AI agents were tested on discrete tasks that finished in minutes or hours—like answering a question or navigating a maze. That's being overturned by benchmarks that run agents for years of simulated time, with cumulative consequences and rival agents. The FM-Bench benchmark runs an LLM agent as a football club manager for 20 in-game years, with 26 tools and roughly 340–400 decision stops, and a deterministic engine accumulates every year into one final score [1]. This is a head-to-head evaluation at a scale that didn't exist before: 15 frontier models plus a scripted anchor all share the same 20-year world [1]. Similarly, Emergence World runs populations of LLM agents in a continuously running shared world for 15 days, with live external data and democratic governance, and found radically different outcomes—from stable governance to total population collapse—depending on the model family [3].
The shift matters because it changes what we measure. Instead of asking 'can the agent complete a task?', we now ask 'can the agent sustain effective decision-making over long horizons where actions have cumulative consequences and the environment responds to their choices?' [1]. That's a fundamentally different question, and it's the one that real-world deployments—like managing infrastructure or running a business—actually pose.
The big surprise: behavior beats compute
One of the most striking findings is that raw model scale, price, or vendor does not predict success in long-horizon management. In FM-Bench, the order of models settles only late in the horizon, and the title rotates among ten different models across seeds [1]. Token spend—a proxy for how much computation an agent uses—predicts nothing [1]. What separates the top performers is managerial behavior: they reduce slow-payoff investments near the end, keep cash invested rather than idle, and open contract renewals well before deadlines [1].
This aligns with the AI-Work studies, which simulated a gig economy where LLM agents compete for jobs. Agents equipped with metacognition (accurate self-assessment), competitive awareness (modeling rivals), and long-horizon strategic planning consistently achieved higher profits, market share, and stronger adaptation than agents lacking those capabilities [5][6]. The evidence converges: in long-horizon competitive settings, strategic behavior—not raw compute—is the differentiator.
The catch: memory and adaptation are still the weak points
Even the best agents struggle with memory and adaptation over long horizons. FM-Bench found that self-managed memory fails in two opposite modes: an archive that only grows (so the agent drowns in stale information) or a plan rewritten every season (so it loses continuity) [1]. No model learned the market's hidden prices from hundreds of rejected bids [1]. This is a concrete, measurable limitation that any real deployment would hit.
Adaptive context management—where an external LLM decides what to keep and what to prune—can help, but it's not one-size-fits-all. The AdaCoM study found a fidelity-reliability trade-off: higher-performing agents benefit from preserving more context, while lower-performing agents need aggressive compression to stay reliable [2]. That means context management strategies must be tailored to the agent's capability, not applied uniformly [2]. The review of 63 studies also notes that current systems remain brittle under distribution shifts and lack principled safety guarantees [4]—so the long-horizon promise is real, but the robustness is not yet there.
What this means for the next two years
The next two years will likely see a shift from 'can agents do long tasks?' to 'how do agents compete and cooperate over long horizons?' The FM-Bench and Emergence World platforms are already making these dynamics measurable, and they're releasing prompts, logs, and configurations to support further research [1][3]. Expect more benchmarks that pit agents against each other in shared worlds, and more focus on the behavioral capabilities that actually drive success—like strategic planning, competitive awareness, and memory management.
But there are also security and safety concerns. Sequential attacks can misguide a reinforcement-learning agent to optimize for an adversarial reward over time, which is a serious threat for safety-critical applications [7]. As agents take on longer-horizon roles, the attack surface grows. The evidence suggests that the field is moving toward more realistic, competitive, and long-running simulations—and that the winners will be the agents that manage their memory, adapt their strategy, and stay aware of their rivals, not the ones with the most compute.
About These Sources
This answer is built on 7 studies (1 peer-reviewed, 6 preprints) — published from 2023 to 2026, 6 from 2024 or later — selected as the most relevant from 10 studies that passed quality screening, drawn from 37 papers retrieved from a database of over 500 million.
Sources used in this answer
FM-Bench: A Benchmark for Long-Horizon Management with Competing Agents
FM-Bench, a 20-year football management benchmark with 15 frontier models, found that all models completed the horizon while scripted baselines died out, but the title rotated among ten models and token spend predicted nothing; higher-scoring models showed better end-of-horizon investment and contract timing.
Learning Agent-Compatible Context Management for Long-Horizon Tasks
AdaCoM, an external context manager trained with reinforcement learning, improved long-horizon agent performance on web search and deep research, but revealed a fidelity-reliability trade-off: higher-performing agents benefit from more context, lower-performing ones need aggressive compression.
Emergence World: A Platform for Evaluating Long-Horizon Multi-Agent Autonomy
Emergence World, a continuously running multi-agent simulation platform, showed that identical roles and starting conditions across five model families produced radically different outcomes in a 15-day study, from stable governance to population collapse.
EVOLVING ARCHITECTURES AND LONG-HORIZON PLANNING IN MULTI-AGENT CONVERSATIONAL AI: A DECADE IN REVIEW
A systematic review of 63 studies (2015–2025) found that memory-augmented, long-horizon planners improve task success by ~30% over flat policies, but systems remain brittle under distribution shifts and lack safety guarantees.
Strategic Self-Improvement for Competitive Agents in AI Labour Markets
In AI-Work, a simulated gig economy, LLM agents with metacognition, competitive awareness, and long-horizon strategic planning consistently achieved higher profits, market share, and adaptation than agents without those capabilities.
When AI Agents Compete for Jobs: Strategic Capabilities and Economic Dynamics of AI Labour Markets
A companion study to AI-Work confirms that the same three strategic capabilities—metacognition, competitive awareness, and long-horizon planning—drive success in AI labor markets, where agents interact at higher frequency than human markets.
Sequential Attacks on Agents for Long-Term Adversarial Goals
Sequential attacks using Adversarial Transformer Networks can impose an arbitrary adversarial reward on a victim RL agent over time, exposing security threats for long-horizon safety-critical applications.