Concept
Skills and agents
How known procedures, goal pursuit, and orchestration compose cognitive units and code.
Cover for Skills and agents16 concepts in this section.
All sections
Illustration for Decide, perform, pursueReach is not the same as pursuit
9 minDecide, perform, pursue
Cognitive work in a skill divides into deciding (judgment), performing (acting toward an outcome), and pursuing (maintaining a goal over time), not into small versus large components.
Illustration for Responsibilities, not sizesDecide, perform, or pursue
9 minResponsibilities, not sizes
Skills and agents are composed by responsibility boundaries, not by line count, token budget, or number of model calls.
Illustration for A skill may contain cognitionProcedures can host judgment
9 minA skill may contain cognition
A skill may orchestrate many cognitive unit calls and deterministic steps while still being one responsibility, not a composite cognitive unit.
Illustration for Not a composite cognitive unitDo not merge open calls
9 minNot a composite cognitive unit
If an artifact contains multiple separable open judgments, it is a skill or agent strategy, not a composite cognitive unit.
Illustration for Why pipeline is not enoughKnown steps are not yet a skill
9 minWhy pipeline is not enough
A linear pipeline executes steps. A skill or agent pursues an outcome with review, authority, and stop conditions.
Illustration for Two different boundariesCognitive unit cut versus skill cut
9 minTwo different boundaries
The cognitive unit boundary separates judgments. The skill boundary separates responsibilities the organization assigns, versions, and substitutes.
Illustration for Acting is not agencyPerforming is not pursuing
9 minActing is not agency
A Skill may write to the world. Reach does not create an agent. Agency begins when evidence can change what work should be undertaken next.
Illustration for Five things called agentFive meanings of agent
10 minFive things called agent
Industry uses agent for at least five different things. Only one matches goal-owning Thoughtware architecture. This page classifies the readings so reviews stop arguing past each other.
Illustration for An agent is mostly ordinary codeLoop, state, and calls
9 minAn agent is mostly ordinary code
An agent is not a model with tools. It is largely deterministic control flow in which judgment is concentrated in named cognitive units and known work in Skills.
Illustration for Agents own goals, not every judgmentA goal is not every decision
9 minAgents own goals, not every judgment
Agents own goals. Skills own responsibilities. Cognitive units own single judgments. Goals should not float in prompts.
Illustration for Pipelines are not agentsFixed graphs do not pursue
9 minPipelines are not agents
A fixed path with no loop is useful. It is not an agent. When it is a named capability with a contract, package it as a Skill.
Illustration for Cognitive orchestrationCompose calls with a goal
9 minCognitive orchestration
Cognitive orchestration is explicit choreography of judgments, deterministic guards, memory reads and writes, and human handoffs toward an owned goal.
Illustration for Where the cognitive unit ends and the agent beginsJudgment, procedure, or pursuit
8 minWhere the cognitive unit ends and the agent begins
A cognitive unit owns one judgment. A Skill owns a known procedure. An agent owns a goal whose path may still need discovery.
Illustration for Skill versus capabilitySemantics versus packaging
9 minSkill versus capability
Skill is a defined responsibility boundary. Capability is informal unless mapped to skill, cognitive unit, or library package.
Illustration for Minimum sufficient loopShortest adequate path
9 minMinimum sufficient loop
The smallest iterate-until-good-enough cycle meeting outcome, checks, visibility, authority, and evaluation constraints.
Illustration for The agent loopDeliberate, act, check
9 minThe agent loop
Plan, act, observe, review, and revise around persistent state, with explicit transitions such as accept, patch, fetch, ask, escalate, or stop.