Latest news with #reinforcementlearning
Finextra
28-07-2025
- Business
- Finextra
Agentic AI in FX: From Automation to Autonomy: By Chandresh Pande
Abstract Imagine a super-smart FX trader on your desk—one who continuously scans global markets, detects macroeconomic shifts, adapts execution strategies in real time, learns from fluctuations, manages risk independently, and spots arbitrage before others even notice. Now picture a back-office specialist who predicts settlement failures, flags reconciliation breaks, updates static data across systems, and ensures regulatory compliance—all without human intervention. Moreover, these guys work 24/7, never complain, take no coffee breaks, and do not even ask for a raise! Too good to be true, right? Think again. What once would have sounded like science fiction is now rapidly becoming reality—powered by Agentic AI: intelligent, autonomous systems that perceive, reason, and act with purpose. Rooted in cognitive science and robotics, agentic systems evolved from early research prototypes into adaptive, autonomous problem-solvers. Advances in reinforcement learning and large language models (LLMs) have enabled agents to make decisions, learn from outcomes, and operate independently in complex domains. With real-time data and scalable computing, finance is emerging as their next frontier. Unlike traditional AI, which passively processes data, agentic systems thrive on feedback loops—observing, deciding, and evolving—making them ideal for the dynamic, high-stakes world of FX trading and operations. With capabilities like autonomous strategy selection, self-directed risk management, and real-time market adaptation, Agentic AI has the potential to transform how institutions engage with FX markets. But as adoption grows, so do questions around oversight, explainability, and trust. Is this the dawn of truly intelligent automation in FX—or just another technological mirage? One thing is clear: the agent is already on the floor. What is Agentic AI? Agentic AI refers to artificial intelligence systems that operate with autonomy, intentionality and adaptability – much like a human agent1. These systems don't just follow pre-defined rules or passively respond to inputs; they set goals, make context aware decisions take actions and learn from the outcomes in a continuous feedback loop. In contrast to traditional models that execute fixed workflows, agentic systems can dynamically change their course based on new information – enabling them to thrive in uncertain, fast changing and fragmented markets like FX markets where milliseconds matter. A major strength of agentic AI lies in multi-agent system (MAS)1, where multiple specialized agents interact and coordinate across different roles, that can be particularly useful in financial systems. In FX environments, MAS can simulate trading, pricing, compliance, risk and settlement via different agents working together to achieve shared goals. This can facilitate simulation and execution of complex workflows like price discovery, order routing, trade matching etc. while also optimizing confirmations, exception handling and settlement workflows in the back office. The distributed nature of MAS improves resiliency, processing speed and enables adaptive response to market. Unlocking Agentic AI in FX The FX market, with its 24x5 trading cycle, deep liquidity, and high volatility, is ideally suited for the integration of agentic AI. These intelligent systems are capable of autonomous decision-making and continuous adaptation, making them valuable in navigating the rapid changes driven by macroeconomic events, geopolitical shifts, and client behaviours. This section highlights a few use-cases how agentic AI can deliver efficiency, reduce risk, and provide strategic advantages across front, middle, and back-office FX functions. These use cases are illustrative not exhaustive. As agentic AI matures, countless other applications will emerge across the FX trade lifecycle. 1. Pre Trade market intelligence and Signal generation Agentic AI systems can autonomously scan and synthesize macroeconomic data, real-time liquidity trends, news feeds, central bank statements, and social media signals. This allows them to generate actionable trade signals or predictive macro views. Additionally, agentic AI can serve as a latency arbitrage hunter by scanning multiple FX trading venues (ECNs, dark pools, etc.) for price discrepancies, where millisecond differences in timing and pricing matter. Example: Prior to an ECB rate decision, an agent might detect tone shifts in ECB speeches and correlate them with historical market reactions. It then feeds these directional insights into the execution algorithm. 2. Autonomous Trade execution These agentic AI systems can use self-evolving execution algorithms that factor in liquidity, order book behaviour, spreads, and volatility in real-time. Unlike static rule-based systems, they dynamically self-tune execution strategies based on objectives such as slippage minimization or speed. Example: An agent detecting a sudden liquidity drop may reroute the order flow or delay execution to prevent slippage, mimicking human trader decision-making but at a machine scale and speed. 3. Liquidity Provision and Market Making Agentic AI systems can operate as autonomous market makers. By monitoring market volatility, client flow, and inventory risks, they can autonomously adjust bid-ask spreads and quote levels. Example: During geo-politically induced volatility, the agent may momentarily widen spreads, then narrow them post-event to restore competitiveness while managing inventory risk. 4. Client behaviour modelling and Personalization These agents can analyze granular client data—such as trading patterns, profitability, and preferences—to segment clients and deliver hyper-personalized strategies. They learn from historical data to forecast behaviour and optimize pricing models or service tiers. Example: A spike in hedging frequency by a client may prompt an alert for the relationship manager to review service models or offer targeted product solutions. 5. Real time Risk monitoring and Response Agentic AI systems can enhance FX risk management by identifying evolving counterparty risks, large directional exposures, or breaches in risk thresholds. They can recommend or auto-execute mitigation actions such as portfolio rebalancing or hedge placement. Example: If an agent detects concentrated exposure due to a correlated client flow, it may autonomously initiate offsetting trades or flag risk teams for pre-emptive action. 6. Settlement failure prediction and intervention Agentic AI can analyze post-trade data across the entire settlement chain to predict which trades are at risk of failing. These agents can use patterns from past settlement failures, counterparty behaviour, payment system data, and real-time exceptions to proactively intervene. They can recommend corrective actions—such as reallocation of funding, client follow-ups, or adjustments in trade instructions—to prevent bottlenecks or penalties. Example: An autonomous 'settlement operations agent' may detect a high probability of failure in a CLS-linked FX leg due to delayed funding from a counterparty, triggering an alert or rebooking logic to avoid settlement disruption. 7. Regulatory Reporting and Compliance monitoring Agentic AI can assist in real-time regulatory compliance by ensuring reporting accuracy across multiple jurisdictions. They automatically validate trade lifecycle data, flag anomalies, and ensure alignment with EMIR, MiFID II, and Dodd-Frank. Example: An AI agent may detect trade discrepancies in timestamps or record-keeping and auto-trigger remediation workflows. Challenges While the potential of agentic AI in financial markets is immense, its safe and effective adoption is fraught with challenges. Below are three critical hurdles that must be addressed before Agentic AI can take the driver's seat in FX world. 1. Autonomy vs. Accountability A core feature of agentic AI is its ability to act autonomously. However, in a highly regulated domain like FX, accountability is paramount. If an autonomous agent executes a trade that results in significant losses or violates regulations, who bears responsibility — the quant who designed the system, the trader who deployed it, or the institution itself? This lack of clarity over responsibility raises serious legal and ethical concerns. Without robust governance structures, auditability, and real-time supervisory frameworks, widespread deployment will remain cautious2. 2. Black Box Behaviour Many agentic AI systems — particularly those leveraging reinforcement learning — behave as 'black boxes,' learning optimal strategies from past data without offering clear rationale for individual decisions. In FX, where compliance and transparency are critical, this opacity is problematic. Regulators increasingly demand explainability and audit trails to justify market behavior. Without transparent decision-making, agentic AI risks introducing systemic vulnerabilities, especially in high-stakes scenarios such as volatility spikes3. 3. Safe Adaptability in Volatile Markets3 Adaptability is one of agentic AI's greatest strengths — but in volatile FX markets, unchecked adaptability can backfire. Constant real-time adjustments to noisy signals can lead to overreactions, unintended feedback loops, or even market destabilization (as seen in past flash crashes4). Rigorous guardrails, staged deployment environments, and stress-testing of agentic behaviors are essential to ensure that 'smart' does not become 'reckless.' The Cutting Edge Leading investment banks are beginning to explore Agentic AI frameworks in controlled environments. JP Morgan5 is leveraging its Athena platform to deploy agent-based systems for risk analytics and trade booking, demonstrating early-stage automation of front office workflows Goldman Sachs5, through its Marquee platform, is employing agents to assist in options pricing and the generation of structured product ideas. Morgan Stanley5 has introduced AskResearchGPT, an agentic model designed to recommend the next best action for trade decisions and to assist in alpha generation, blending research automation with trading insight. Citi5 is utilizing agentic AI in FX for both market making and smart order routing within the fragmented FX markets, showcasing a move towards autonomous execution and adaptive flow management. Two Sigma's1 Venn platform combines market analytics with reinforcement learning agents to dynamically calibrate investment strategies based on changing market conditions. JP Morgan's1 LOXM system, which integrates agentic AI to analyze market data, news, and social media, uncovers real-time investment opportunities. These initiatives signal a growing institutional appetite to harness agentic AI not just for efficiency, but for a strategic edge — driving a shift from static automation to autonomous, intelligent financial systems. Conclusion Agentic AI marks a significant leap in the evolution of financial automation—shifting from passive tools to autonomous, goal-oriented digital agents capable of executing complex decisions across the FX trade lifecycle. As illustrated at the beginning of this article through the imagined trader and operations personas, these agents are no longer confined to generating insights; they actively trade, reconcile, hedge, and adapt—continuously learning from their environment to meet strategic objectives. The use cases across the front, middle, and back office are compelling: autonomous execution, arbitrage detection, proactive risk mitigation, dynamic margin management, and intelligent exception handling. Each demonstrates how agentic AI can reshape FX workflows with speed, precision, and round-the-clock responsiveness. Yet, these possibilities come with real challenges. From autonomy vs. accountability to the opacity of black-box decision-making and the risk of unintended feedback loops in live trading environments, the path to widespread adoption must be tread with caution and clarity. Agentic systems must be deployed with human oversight, robust guardrails, and explainability built in from day one. We also see that leading investment banks and financial firms are exploring the possibilities, but these are still in early stages. Some are piloting "trading copilots" that work alongside human dealers; others are experimenting with agentic systems for post-trade workflows. These early initiatives signal both interest and caution—a recognition that agentic systems can bring scale and intelligence, but only when aligned with enterprise goals, operational resilience, and regulatory trust. Ultimately, the future of FX will not be human or machine—but human and machine, working in tandem. Agentic AI won't replace traders or operations teams but will act as tireless digital teammates, amplifying capabilities, enhancing decision-making, and navigating the increasingly complex FX landscape with intelligence, autonomy, and precision. References 1. 'Building Agentic AI Systems' by Anjanava Biswas & Wrick Talukdar, Packt Publishing. 2. Gasser, U., & Almeida, V. A. (2017). "A Layered Model for AI Governance." Harvard Journal of Law & Technology. 3. European Securities and Markets Authority (2022). 'Final Report: Guidelines on AI in Financial Markets.' 4. Kirilenko, A. A., Kyle, A. S., Samadi, M., & Tuzun, T. (2017). 'The Flash Crash: The Impact of High Frequency Trading on an Electronic Market.' Journal of Finance. 5. 6. 7.
WIRED
16-07-2025
- Business
- WIRED
Another High-Profile OpenAI Researcher Departs for Meta
Jul 15, 2025 10:56 PM Jason Wei, who worked on OpenAI's o3 and deep research models, will be joining Meta's superintelligence lab. His colleague Hyung Won Chung is also joining Meta, a source tells WIRED. ILLUSTRATION: WIRED STAFF; GETTY IMAGES OpenAI researcher Jason Wei is joining Meta's new superintelligence lab, according to multiple sources familiar with the matter. Wei worked on OpenAI's o3 and deep research models, according to his personal website. He joined OpenAI in 2023 after a stint at Google, where he worked on chain-of-thought research, which involves training an AI model to process complex queries step-by-step. At OpenAI, Wei became a self-described 'diehard' for reinforcement learning, a method of training or refining an AI model with positive or negative feedback. It's become a promising area of AI research—one that several of the researchers Meta has hired for its superintelligence team specialize in. One source tells WIRED that another OpenAI researcher, Hyung Won Chung, will also be joining Meta. Multiple sources confirm that both Wei and Chung's internal OpenAI Slack profiles are currently deactivated. OpenAI, Meta, Wei, and Chung did not immediately respond to requests for comment from WIRED. Chung worked on some of the same projects at OpenAI as Wei, including deep research and OpenAI's o1 model, according to Chung's personal website. His research is primarily focused on reasoning and agents, the website says. Chung overlapped with Wei at Google as well, and joined OpenAI at the same time as Wei, per their LinkedIn profiles. Multiple sources tell WIRED that Wei and Chung have a close working relationship. Meta has previously poached groups of researchers that have experience working together for its new superintelligence lab, including a trio from OpenAI's Switzerland office that joined the ChatGPT maker from Google. Meta has been going on a poaching spree over the past month, offering up to $300 million over four years to top AI talent. WIRED reported late last month that Meta CEO Mark Zuckerberg sent an internal memo to staff that laid out a fresh plan for the company's AI efforts. It included a list of new staffers for the superintelligence team, most of whom had been recruited from OpenAI. The hiring frenzy shows no signs of slowing down, and OpenAI has been fighting back. Just last week, WIRED reported that OpenAI had recruited four high-ranking engineers from Tesla, xAI, and Meta. On Tuesday, Wei shared a post on social media reflecting on what he called 'an important lesson' that reinforcement learning taught him 'about how to live my own life.' In life, (and when building AI models), imitation is good and you have to do it at first, Wei wrote. But 'beating the teacher requires walking your own path and taking risks and rewards from the environment.'
Geek Wire
08-07-2025
- Business
- Geek Wire
Pokee AI, a new AI agent startup led by ex-Meta manager, lands $12M to automate online workflows
GeekWire's startup coverage documents the Pacific Northwest entrepreneurial scene. Sign up for our weekly startup newsletter , and check out the GeekWire funding tracker and venture capital directory . From left: Michael Cai, head of product engineering), Christopher Wu (head of ML engineering), Zheqing (Bill) Zhu (founder and CEO), and Yi Wan (founding research scientist). (Pokee AI Photo) Point72 Ventures led a $12 million seed round for Pokee AI, a new Seattle-area startup aiming to build AI agents that automate online workflows. Qualcomm and Samsung also invested. Founded last year by a former Meta manager and Stanford Ph.D., Pokee AI is developing a general-purpose AI agent designed to plan, reason, and take actions across thousands of internet tools and platforms — everything from generating marketing videos to formatting slides and posting on social media. Pokee is one of many new startups building so-called AI agents, or autonomous systems that execute tasks beyond static outputs like text or code. These companies are attracting plenty of attention from early stage investors. Pokee differentiates itself by applying reinforcement learning to help agents sequence and use tools efficiently, rather than relying solely on large language models. Reinforcement learning is a type of machine learning where an agent takes actions and learns through trial and error. The company claims its technology delivers 'over 97% accuracy when selecting from thousands of tools,' and avoids the limitations of function-calling via LLMs by offloading planning to a custom-built AI agent. 'The AI world has solved the generation problem, but no one has solved the execution problem,' said CEO Zheqing (Bill) Zhu. 'We want to solve the execution problem.' One early use case is social media marketing, where the agent can create content, enhance media, post across platforms, and monitor engagement. Pokee is not generating revenue but has design partners and is working with Google on enterprise partnerships. Its platform is now in public beta. Pokee supports integrations with Google Workspace, Meta, LinkedIn, YouTube, Jira, GitHub, Slack, Notion, and other popular web services. It eventually plans to release tools for enterprise users. Zhu was previously head of applied reinforcement learning at Meta, where he worked for more than seven years. He completed his Ph.D. at Stanford in the same field and did his graduate and undergrad studies at Duke. Pokee's other co-founders include Michael Cai, head of product engineering; Christopher Wu, head of ML engineering; and Yi Wan, founding research scientist. Other backers include Salience Capital, SCB 10X, Typeface founder Abhay Parasnis, and former Intel board member Lip-bu Tan. The company has around 10 employees and is hiring for open roles in Bellevue and the Bay Area.
News.com.au
04-06-2025
- Science
- News.com.au
AI robot developed to play badminton
A Swiss-led team has developed an AI-legged robot which can play badminton against humans. The Robot uses reinforcement learning, where it learns by trying different actions to make better decisions. The AI utilises vision, movement, arm control, and a perception noise model trained on real-world camera data to maintain consistent performance. It can accurately predict a shuttlecock's trajectory and navigate the game area.
Forbes
23-05-2025
- Forbes
Agentic AI: Winning In A World That Doesn't Work That Way
Agentic AI is being trained to 'win.' But human systems aren't games—they're stories. The consequences of confusing the two will define the next decade. Agentic AI is being built on the assumption that the world is a game—one where every decision can be parsed into players, strategies, outcomes, and a final payoff. This isn't a metaphor. It's code. In multi-agent reinforcement learning (MARL), agents use Q-functions to estimate the value of actions in a given state to converge toward an optimal policy. MARL underpins many of today's agentic systems. A Q-function is a mathematical model that tells an AI agent how valuable a particular action is in a given context—essentially, it's a way of learning what to do and when to maximize long-term reward. But 'optimal' depends entirely on the game's structure—what's rewarded and penalized and what constitutes 'success.' Q-learning becomes a hall of mirrors when the world isn't a game. Optimization spirals into exploitation. MARL becomes even more hazardous because agents must not only learn their policies but also anticipate the strategies of others, often in adversarial or rapidly shifting contexts, as seen in systems like OpenAI Five or AlphaStar. At the heart of agentic AI—AI designed to act autonomously—is a set of training systems built on game theory: multi-agent reinforcement learning, adversarial modeling, and competitive optimization. While tools like ChatGPT generate content based on probability and pattern-matching, agentic AI systems are being built to make autonomous decisions and pursue goals—a shift that dramatically raises both potential and risk. The problem is that human life doesn't (and more importantly, shouldn't be induced to) work that way. Game theory is a powerful tool for analyzing structured interactions, such as poker, price wars, and Cold War standoffs. Those are not games. They are stories. And storytelling isn't ornamental—it's structural. We are, as many have argued, not just homo sapiens but homo narrans: the storytelling species. Through narrative, we encode memory, make meaning, extend trust, and shape identity. Stories aren't how we escape uncertainty—they're how we navigate it. They are the bridge between information and action, between fact and value. To train machines to optimize for narrow wins inside rigid systems is to ignore the central mechanism by which humans survive uncertainty: We don't game the future—we narrate our way through it. And training agents to 'win' in an environment with no final state isn't just shortsighted—it's dangerous. Game theory assumes a closed loop: Simon Sinek famously argued that business is an 'infinite game.' But agentic AI doesn't play infinite games—it optimizes finite simulations. The result is a system with power and speed, but lacking intuition for context collapse. Even John Nash, the father of equilibrium theory, understood its fragility. His later work acknowledged that real-life decision-making is warped by psychology, asymmetry, and noise. We've ignored that nuance. But in real life—especially in business—the players change, the rules mutate, and the payoffs are subjective. Even worse, the goals themselves evolve mid-game. In AI development, reinforcement learning doesn't account for that. It doesn't handle shifting values. It handles reward functions. So, we get agents trained to pursue narrow, static goals in an inherently fluid and relational environment. That's how you get emergent failures—agents exploiting loopholes, corrupting signals, or spiraling into self-reinforcing error loops. We're not teaching AI to think. We're teaching it to compete in a hallucinated tournament. This is the crux: humans are not rational players in closed systems. We don't maximize. We mythologize. Evolution doesn't optimize like machines do—it tolerates failure, ambiguity, and irrationality as long as the species endures. It is selected not just for survival and cooperation but also for story-making because narrative is how humans make sense of uncertainty. People don't start companies or empires solely to 'win.' We often do it to be remembered. We blow up careers to protect pride. We endure pain to fulfill prophecy. These are not strategies—they're spiritual motivations. And they're either illegible or invisible to machine learning systems that see the world as a closed loop of inputs and rewards. We pursue status, signal loyalty, perform identity, and court ruin—sometimes on purpose. You can simulate 'greed' or 'dominance' by tweaking rewards, but these are surface-level proxies. As Stuart Russell notes, the appearance of intent is not intent. Machines do not want—they merely weigh. When agents start interacting under misaligned or rigid utility functions, the system doesn't stabilize. It fractures. Inter-agent error cascades, opaque communications, and emerging instability are the hallmarks of agents trying to navigate a reality they were never built to understand. Imagine a patient sitting across from a doctor with a series of ambiguous symptoms—fatigue, brain fog, and minor chest pain. The patient has a family history of heart disease, but their test results are technically 'within range.' Nothing triggers a hard diagnostic threshold. An AI assistant, trained on thousands of cases and reward-optimized for diagnostic accuracy, might suggest no immediate intervention—maybe recommend sleep, hydration, and follow-up in six months. The physician, though, hesitates. Not because of data but because of tone, posture, and eye contact, because the patient reminds them of someone, because something feels off, even though it doesn't compute. So, the doctor ordered the CT scan against the algorithm's advice. They find the early-stage arterial blockage. They save the patient's life. Why did the doctor do it? Not because the model predicted it. Because humans don't operate on probability alone—we act on a hunch, harm avoidance, pattern distortion, and story. We're trained not only to optimize for outcomes but also to prevent regret. A system trained to 'win' would have scored itself ideally. It followed the rules. But perfect logic in an imperfect world doesn't make you intelligent—it makes you brittle. The fundamental flaw in agentic AI isn't technical—it's conceptual. It's not that the systems don't work; they're working with the wrong metaphor. We didn't build these agents to think. We built them to play. We didn't build agents for reality. We built them for legibility. Game theory became the scaffolding because it provided a structure, offering bounded rules, rational actors, and defined outcomes. It gave engineers something clean to optimize. But intelligence doesn't emerge from structure; it arises from adaptation within disorder. The gamification of our informational matrix isn't neutral. It's an ideological architecture that recodes ambiguity as inefficiency and remaps agency into pre-scored behavior. This isn't just a technical concern—it's an ethical one. As AI systems embed values through design, the question becomes: whose values? In the wild, intelligence isn't about winning. It's about not disappearing. It's about adjusting your behavior when the ground shifts under you because it will. No perfect endgames exist in nature, business, politics, and human relationships; they are just survivable next steps. Agentic AI, trained on games, expects clarity. But the world doesn't offer clarity. It offers pressure. And pressure doesn't reward precision—it rewards persistence. This is the failure point. We're asking machines to act intelligently inside a metaphor never built to explain real life. We simulate cognition in a sandbox while the storm rages outside its walls. If we want beneficial machines, we need to abandon the myth of the game and embrace the truth of the environment: open systems, shifting players, evolving values. Intelligence isn't about control. It's about adjustment, not the ability to dominate, but the ability to remain. While we continue to build synthetic minds to win fictional games, the actual value surfaces elsewhere: in machines that don't need to want. They need to move. Mechanized labor—autonomous systems in logistics, agriculture, manufacturing, and defense—isn't trying to win anything. It's trying to function. To survive conditions. To optimize inputs into physical output. There's no illusion of consciousness—just a cold, perfect feedback loop between action and outcome. Unlike synthetic cognition, mechanized labor solves problems the market understands: how to scale without hiring, operate in unstable environments, and cut carbon and cost simultaneously. Companies like John Deere are already deploying autonomous tractors that don't need roads or road signs. Amazon has doubled its robotics fleet in three years. These machines aren't trying to win. They're trying not to break. And that's why capital is quietly pouring into it. The next trillion-dollar boom won't be in artificial general intelligence. It'll be in autonomous physicality. The platforms we think of as background are about to become intelligent actors in their own right. 'We have become tools of our tools,' wrote Thoreau in 'Walden' in 1854, just when the industrial revolution began to transform not just Concord, but America, Europe, and the world. Intriguingly, Thoreau includes mortgage and rent as 'modern tools' to which we voluntarily enslave ourselves. What Thoreau was pointing to with his experiment in the woods was how our infrastructure, the material conditions of our existence, comes to seem to us 'natural' and inevitable, and that we may be sacrificing more than we realize to maintain that infrastructure. AI - intelligent, autonomous tools - represents a categorical shift in how we coexist with our infrastructure. Infrastructure isn't just how we move people, goods, and data. It's no longer just pipes, power, and signals. It's 'thinking' now—processing, predicting, even deciding on our behalf. What was once physical has fused with the informational. The external world and our internal systems of meaning are no longer separate. That merger isn't just technical—it's existential. And the implications? We're not ready. But if AI is to become all of our closest, most intimate companions, we should be clear on what it is, exactly, that we have trained it, and allowed it, to do. This isn't just logistics. It's the emergence of an industrial nervous system. And it doesn't need to 'win.' It needs to scale, persist, and adapt—without narrative. We're building agentic AI to simulate our most performative instincts while ignoring our most fundamental one: persistence. The world isn't a game. It's a fluid network of shifting players, incomplete information, and evolving values. To train machines as if it's a fixed competition is to misunderstand the world and ourselves. We are increasingly deputizing machines to answer questions we haven't finished asking, shaping a world that feels more like croquet with the Queen of Hearts in Alice's Adventures in Wonderland: a game rigged in advance, played for stakes we don't fully understand. If intelligence is defined by adaptability, not perfection, endurance becomes the ultimate metric. What persists shapes. What bends survives. We don't need machines that solve perfect problems. We need machines that function under imperfect truths. The future isn't about agentic AI that beats us at games we made up. It's about agentic AI that can operate in the parts of the world we barely understand—but still depend on.
