AI deployments that perceive an environment, set sub-goals, and execute multi-step actions autonomously toward an objective, with limited human intervention between decision and execution. In finance they span agentic trading, research, and operations — distinguished from generative AI by their capacity to act rather than merely produce outputs.
How it works
An agentic system couples a reasoning model (often an LLM) with tools, memory, and a feedback loop: it decomposes a goal into tasks, calls APIs or execution venues, observes results, and iterates without per-step human sign-off. Autonomy sits on a spectrum from human-in-the-loop approval to fully unsupervised action, which is the axis that matters for risk.
Why it matters now
Regulators including the BIS, FSB, and IOSCO are flagging financial-stability risks from agentic systems with limited human oversight — correlated behaviour from shared models, herding, and a widened attack surface — making the autonomy/oversight trade-off a live 2025-2026 supervisory question.
Example
If multiple trading desks deploy agents built on the same foundation model and similar prompts, a common signal could trigger synchronised liquidation — an algorithmic flash-crash analogue. The May 2010 flash crash, driven by automated but non-agentic order routing, is the canonical cautionary precedent; agentic autonomy raises the stakes by removing the human pause between decision and execution.
Frequently asked
- What are agentic systems in finance?
- Agentic systems are AI deployments that perceive an environment, set sub-goals, and execute multi-step actions autonomously toward an objective with limited human intervention. In finance they span agentic trading
Glossary · agentic systems / agentic trading
Agentic systems are AI-driven architectures in which software agents trade, lend, allocate, and execute financial decisions with limited human oversight, acting autonomously within delegated mandates. Agentic trading applies this to markets: agents place orders, manage risk, and rebalance portfolios on behalf of users rather than merely surfacing recommendations.
Read more →
- How do agentic systems differ from generative AI?
- Agentic systems act; generative AI produces. A generative model returns an output — text, code, a forecast — and stops. An agentic system couples that model with tools, memory, and a feedback loop, decomposing a goal into tasks, calling execution venues, observing results, and iterating without per-step human sign-off. The defining axis is autonomy: whether a human pause sits between decision and execution.
- Why do regulators worry about agentic systems?
- Regulators worry that agentic systems with limited human oversight create new financial-stability risks. The BIS, FSB, and IOSCO have flagged correlated behaviour from shared foundation models, herding, and a widened attack surface during 2025–2026. Because agents act without a human pause between decision and execution, errors or shared signals can propagate faster than supervisors or circuit-breakers can intervene.
- Can agentic trading systems cause a flash crash?
- Agentic trading systems can plausibly amplify flash-crash dynamics by removing the human pause between decision and execution. If multiple desks deploy agents on the same foundation model with similar prompts, a common signal could trigger synchronised liquidation. The May 2010 flash crash — driven by automated but non-agentic order routing — is the canonical precedent that agentic autonomy makes more acute.
- What does the autonomy spectrum mean for agentic AI risk?
- The autonomy spectrum runs from human-in-the-loop approval to fully unsupervised action, and it is the single axis that governs agentic AI risk. At the supervised end, a human approves each action; at the autonomous end, agents execute multi-step plans alone. Greater autonomy raises throughput and removes the human checkpoint, which is precisely the trade-off supervisors are scrutinising in 2025–2026.