AI agent
Software that selects and carries out actions toward a goal using inputs, tools and a control loop. Autonomy varies by implementation.
Watch for: Calling something an agent does not establish its permissions or competence.
DN FIELD GUIDE · INITIAL RELEASE
DeFi & Agentic Finance Dictionary
Learn 36 foundational concepts through plain definitions, worked scenarios and failure modes. Each concept has a stable page and links to primary documentation.
Definitions and illustrative examples are DN editorial explanations. Source links provide technical context; they are not endorsements of these examples or a universal taxonomy.
36 results
Agentic finance
Financial workflows where software agents help plan, monitor or execute actions under delegated authority.
Watch for: Automating a decision does not remove financial, legal or operational responsibility.
Agent harness
The surrounding system that controls tool access, execution, memory and the lifecycle of an agent.
Watch for: Model quality cannot compensate for unrestricted execution permissions.
Tool calling
A model emits a structured request for an external function; the application validates and executes it.
Watch for: A model request is not itself user consent to execute a transaction.
Workflow
A defined sequence of tasks with explicit transitions and conditions.
Watch for: A fixed workflow and an open-ended agent have different failure boundaries.
Human in the loop
A control point requiring a person to review or intervene in a workflow.
Watch for: A vague approval prompt can hide material changes in amount or destination.
Checkpoint
A saved execution state used to resume or inspect a workflow.
Watch for: Resuming old state can reuse outdated market data or approvals.
Model Context Protocol (MCP)
A protocol connecting applications to tools, resources and contextual data through servers and clients.
Watch for: A discoverable server is not automatically trustworthy or authorised.
Agent-to-Agent (A2A)
A protocol for communication and task coordination between agent systems.
Watch for: Interoperability does not prove the other agent is honest or competent.
Agent card
A discovery document describing an agent service and its advertised capabilities.
Watch for: Advertised capabilities are claims; authenticate the service separately.
Prompt injection
Untrusted content attempts to redirect an AI system away from its intended instructions.
Watch for: Treat retrieved documents as data, not new authority.
Least privilege
Granting only the permissions needed for a particular task.
Watch for: An API key with withdrawal rights can expose funds even if a bot only intends to trade.
Permission scope
The explicit set of actions or resources an authorisation permits.
Watch for: A broad scope increases the impact of compromise.
Token passthrough
Forwarding a received access token to a downstream service without the appropriate audience boundary.
Watch for: This can bypass controls and obscure accountability; MCP forbids this pattern.
Server-side request forgery (SSRF)
A service is induced to fetch a destination it should not access.
Watch for: Unrestricted URL fetching can expose internal systems and credentials.
Delegated authority
Permission for another actor or program to perform a defined action on behalf of a principal.
Watch for: Delegation needs bounds, expiry and an effective revocation route.
Smart contract
Code deployed to a blockchain that executes according to the network’s rules.
Watch for: Code can contain bugs or rely on privileged administrators.
Wallet
Software or hardware used to manage keys and interact with blockchain accounts.
Watch for: The interface may conceal what a signature actually authorises.
Private key
Secret cryptographic material used to authorise actions for an account.
Watch for: Key possession does not prove an action reflected the owner’s intent.
Account abstraction
Account designs that permit programmable validation and execution rules beyond a simple externally owned account.
Watch for: Custom validation and upgrade controls add their own implementation risks.
Token approval
An allowance that permits a spender to transfer tokens within a specified limit.
Watch for: An unlimited allowance can outlive the intended transaction.
Oracle
A mechanism supplying external information to a blockchain application.
Watch for: Bad or stale information can trigger incorrect financial actions.
Gas
A unit measuring computation and execution work on networks such as Ethereum.
Watch for: The cost depends on the network and transaction; a failed execution can still cost gas.
Finality
The degree of assurance that a recorded transaction will not be reversed under the network’s rules.
Watch for: A transaction appearing in a block is not always sufficient settlement assurance.
Maximal extractable value (MEV)
Value arising from including, excluding or ordering transactions within block production.
Watch for: Quoted output can differ from the eventual execution outcome.
Slippage
The difference between a reference or expected execution price and the achieved price.
Watch for: A slippage tolerance limits acceptable execution conditions; it does not guarantee a fill.
Bid-ask spread
The gap between the best displayed buying and selling prices for a market at a moment in time.
Watch for: Zero explicit fees do not imply zero execution cost.
Time-weighted average price (TWAP)
An execution approach that divides an order across time rather than submitting its entire amount at once.
Watch for: Market movement, outages and shallow liquidity can still produce poor fills.
Limit order
An instruction to trade only at a specified price or better.
Watch for: Price protection does not guarantee execution.
Stop order
An order condition that activates an execution instruction when a specified trigger is reached.
Watch for: Triggering is not the same as filling, especially during fast markets.
Backtest
A historical simulation used to examine how a strategy would have behaved under specified assumptions.
Watch for: Look-ahead bias, overfitting and omitted costs can make results misleading.
Paper trading
A simulation of trading without putting live capital at risk.
Watch for: Simulated fills do not establish live execution quality.
Idempotency
A design property where repeating a request does not create an additional unintended effect.
Watch for: Retries without duplicate protection can send funds twice.
Agent evaluation
A repeatable process for measuring behaviour against defined tasks and failure criteria.
Watch for: A successful demo is not a broad reliability benchmark.
Risk budget
A defined limit on the financial or operational exposure a workflow may create.
Watch for: A per-order cap alone may permit many repeated orders.
Kill switch
A mechanism intended to halt further automated action.
Watch for: Stopping software may not cancel existing orders or reverse completed transfers.
Evidence before authority
DN’s reference hub connects concepts, official sources and practical review questions. It is an editorial reference, not a security certification or legal advice. Entries are dated snapshots, not a continuously monitored feed. Coverage expands only after review.
Commercial relationships do not determine definitions, regulatory conclusions or default tool order. Follow provider eligibility rules and verify permissions before execution.
Suggest a correction · Dictionary JSON · Regulation JSON · Tools JSON