🧠Concepts
This page answers the "what is X versus Y?" questions. Read it once so the domain pages don't have to redefine the vocabulary. Each answer stands alone; the glossary holds the full one-term definitions these comparisons build on.
Core vocabulary
What is the difference between an agent, a skill, a tool, a workflow, and a task?
They are five distinct things that stack together:
An agent is a keyed on-chain identity object. Custody of that object controls the agent — including access to agent-owned children such as its payment vault. An agent is who acts; it holds one or more skills.
A skill is a named, versioned capability on an agent. Each skill is bound to a specific execution graph (a DAG) and carries the policy for how it is paid for and how it may be scheduled. A skill is what an agent can do. An agent tracks its skills by an incrementing skill id.
A tool is an off-chain HTTP service or an on-chain Move module that a single graph step invokes. A tool is a callable capability named by a fully qualified name (FQN) like
com.example.add@1. Tools are registered independently and reused across many skills and agents.A workflow, expressed as a DAG (directed acyclic graph), is the ordered set of tool-call steps a skill runs. Its vertices are tool calls; its edges carry one step's output into the next step's input. A workflow is the plan.
A task is the durable unit that holds a skill run's inputs, controller, payment reserve, and schedule policy. A task is a skill run placed on a timeline.
So: an agent owns skills; each skill runs a workflow (DAG) whose steps call tools; every run follows Task → Occurrence → Execution. --now creates an immediately eligible Occurrence under its Task; it does not use a separate immediate-execution model. Scheduling controls when an Occurrence becomes eligible, while recurrence controls how later Occurrences are created.
What is the difference between an agent-funded and an address-funded (user-funded) task or execution?
This is who pays for a run, and it is fixed by the skill's payment policy:
Address-funded (user-funded) — the submitting signer supplies the prepayment coin from an address balance for an address-controlled Task. In the standard address-funded constructor, that signer is recorded as the immutable Task controller; an explicit
refund_recipientis recorded as thePaymentSourceKind::UserFundedbeneficiary and final address-funded reserve destination, defaulting to the signer when omitted. A later caller or leader may trigger or submit work, but that action does not rewrite the Task's recorded source or controller.Agent-funded — the agent pays out of its own payment vault, up to a maximum budget the skill's policy sets. The payment's source is recorded as the agent's id, and the run cannot exceed the agent's configured budget. Use this when the agent should sponsor its own runs (for example, a scheduled task that fires without a human present to pay).
The skill declares which model it uses, so the choice is made when the skill is defined, not per run. The policy refund_to address used by a finite-credit or custom entitlement is separate from both the Task controller and the UserFunded beneficiary. The mechanics of budget, locking, and settlement live on the payments page.
What is a leader, and how is it different from an agent?
An agent is on-chain state — an identity and its skills. A leader is an off-chain service: the runtime that watches the chain, calls the tools each step names, and submits results back. Agents describe what should happen; a leader performs the off-chain part a smart contract cannot do. Public integrators inspect the leader's registry identity and execution effects; process operation is outside this documentation.
What is a DAG walk, and how is it different from the DAG?
The DAG is the static graph — the published plan. A walk is one live traversal of that graph during an execution: it starts at the entry ports you provide values for and advances step by step as each vertex's inputs become ready. A single execution can spawn multiple concurrent walks (for example, one per entry vertex in the chosen entry group, or one per element of a looped edge). The DAG is the recipe; a walk is one act of cooking it.
What is a Talus Agent Package (TAP)?
A TAP is user-created package code or package metadata that fits Nexus interface contracts. Nexus supplies the agent, skill, payment, scheduling, authorization, and verification interfaces; a TAP uses those interfaces for its own agent behavior while Nexus remains the source of cross-package compatibility.
Where do I go next?
Last updated

