Concepts
All concepts
67 concepts organized in six layers. Each with one sentence, an example, the common mistake, how to apply it and when it’s done.
67 / 67
Agent Execution
Agent
Receives a goal, decides on actions, uses tools, observes, and iterates. A chatbot only responds.
Agent Execution
Agent Teams
The org chart applied to agents.
Agent Execution
APIs
A Cognitive Enterprise treats "having an API" as a prerequisite for any new system.
Agent Execution
CLI
The gateway through which a program is operated by text, without a screen. It’s what separates software that can be delegated from software that only accepts a human hand.
Agent Execution
Context Engineering
It is to context what capital management is to cash.
Agent Execution
Context Window
The scarcest resource in agentic operations.
Agent Execution
Context7
Up-to-date library documentation directly in the agent’s context.
Agent Execution
File System Access
What enables persistent state and files as shared memory between agents.
Agent Execution
Handoff / Escalation
Clear escalation criteria are what prevent an agent from getting stuck or acting beyond its authority.
Agent Execution
Harness
The model is the engine; the harness is the car.
Agent Execution
Hooks
Automate governance without relying on the model to remember.
Agent Execution
Knobs
The system’s tuning knobs: explicit parameters instead of fixed behavior.
Agent Execution
LLM
The engine. It receives text and predicts the most likely continuation. Memory, decision-making, and action are built around it.
Agent Execution
Loop
Think → act → observe → repeat. Every agentic architecture is a loop with controls around it.
Agent Execution
MCP (Model Context Protocol)
The USB-C of the Cognitive Enterprise stack.
Agent Execution
Memory
Without memory, every session starts from scratch.
Agent Execution
Multi-agent
Only worth it when a single agent can’t handle the job, because coordination has a cost.
Agent Execution
Orchestrator / Routing
The traffic manager: expensive model for difficult tasks, cheaper model for triage.
Agent Execution
Python
The lingua franca that agents write best and the default format for verifiable automation.
Agent Execution
Ralph Loop
For long tasks: the agent wakes up, reads the state, makes some progress, saves, and starts over.
Agent Execution
Sandbox
Prerequisite for granting autonomy safely: errors in the sandbox do not leak into production.
Agent Execution
Skills
An executable SOP: folders with instructions, examples, and scripts loaded on demand.
Agent Execution
Subagent
Child agent with its own context: parallelizes without polluting the main agent’s context.
Agent Execution
Token
The piece of text the model counts. It is the billing unit and what fills the context window, but never your unit of outcome.
Agent Execution
Tools
Without tools, the model only talks. With tools, it gets things done.
Agent Execution
Webhooks
The inverse of the API: they enable agents that react to events, not just ones triggered by humans.
Engineering Process
Compound Engineering
The unit of value is not the completed task; it is accumulated capability.
Engineering Process
Design Mock
The system represented visually before implementation, so the agent does not invent the interface.
Engineering Process
Research
Separating investigation from execution prevents the agent from building on a wrong assumption.
Engineering Process
Rollback / Idempotency
Every action is classified as reversible or irreversible, and irreversible ones require a gate.
Engineering Process
Self-verification
The agent checks its own work before delivering it and leaves evidence that it checked.
Engineering Process
Spec / PRD-driven development
Humans write specs; agents write the implementation. The spec becomes the contract.
Engineering Process
State Management / Checkpointing
It’s what makes long tasks robust instead of fragile.
Engineering Process
TDD (Test-Driven Development)
In an agentic context, the test becomes the executable specification the agent uses to know whether it got it right.
Engineering Process
Test Sprint
A short, focused testing cycle, run by agents in parallel, before moving forward.
Evaluation and Reliability
Determinism vs. Variance
The same input can generate different outputs. The right question isn’t “does it vary?” It’s “where can it vary?”
Evaluation and Reliability
Evals
Test cases plus objective criteria, run after every change. Without evals, every prompt tweak is guesswork.
Evaluation and Reliability
Golden Dataset
The operation’s answer key: examples with a known correct answer, deliberately including the hard cases.
Evaluation and Reliability
Human baseline
How accurate the process is today when done by people. Without this number, "the agent made a mistake" means nothing.
Evaluation and Reliability
LLM-as-Judge
One model evaluates another model’s output at scale against criteria that only a human would normally judge, as long as it has a rubric and calibration.
Evaluation and Reliability
Monitoring (Traces)
Eval happens before, in the lab. Monitoring happens after, in production. You need both.
Evaluation and Reliability
Regression
A change breaks something that used to work. In prompt systems, this happens silently.
Evaluation and Reliability
Satisfaction (% achieved)
The informal version of the eval. Any criterion that repeats should become an item in the suite.
Governance and Control
Audit Trail / Provenance
The legal-organizational version of the technical trace.
Governance and Control
Guardrails
Hard constraints implemented in code, not in the prompt.
Governance and Control
Human-in-the-Loop / Approval Gates
Defining where the gates sit is the organizational design of the Cognitive Enterprise.
Governance and Control
Kill Switch / Circuit Breaker
How to shut down an agent when something goes wrong, and the automatic version, which stops on its own.
Governance and Control
Least Privilege / Permissioning
Marketing agent does not read payroll.
Governance and Control
Levels of Autonomy
1 reads, 2 suggests, 3 acts with approval, 4 acts on its own. Moves up a level when evals prove it.
Governance and Control
Policies / Business Rules
They need to exist in machine-readable format to become guardrails.
Governance and Control
Prompt Injection
External content with malicious instructions that the agent mistakes for a legitimate command.
Knowledge and Semantics
Graph Engineering
Store company knowledge as things and relationships, not as loose chunks of text. This is what makes it possible to answer a question that spans multiple records.
Knowledge and Semantics
Ontology
The formal map of what the company calls a customer, lead, and order, and how they connect.
Knowledge and Semantics
RAG / Retrieval
How institutional knowledge becomes input for the agent without retraining the model.
Knowledge and Semantics
SOP-as-Code
Documentation stops describing the work and becomes the work.
Organizational Design and Economics
Agent Unit Economics
Cost per token → cost per task → cost per outcome. Without that calculation, you cannot decide what to automate.
Organizational Design and Economics
AI Factory
The company as a factory: the input is context, the output is an executed decision.
Organizational Design and Economics
Cognitive Enterprise
A company designed from the ground up with agents as the primary workforce.
Organizational Design and Economics
DRI (Directly Responsible Individual)
The DRI manages agents, and accountability still has a first and last name.
Organizational Design and Economics
Goal
The verifiable objective given to the agent for a task. It is different from the outcome, and the system breaks when the two become misaligned.
Organizational Design and Economics
Internal Transaction Cost
The economic thesis of the Cognitive Enterprise: agents collapse the cost of internal coordination.
Organizational Design and Economics
Process
Work that repeats: it starts with a trigger, passes through hands, ends in an outcome. It’s what gets sliced into tasks.
Organizational Design and Economics
Requisite Variety (Ashby)
The control system needs as much variety as the environment it controls; otherwise, it fails on exceptions.
Organizational Design and Economics
Result / Outcome
What the company wants in the world. Every design starts here, before any agent.
Organizational Design and Economics
Span of Control (redefined)
How many agents a human can supervise effectively. It depends on evals, traces, and gates.
Organizational Design and Economics
Task
The unit of work that is delegated as a whole. Getting its size wrong is the project’s most expensive mistake.
Organizational Design and Economics
VSM (Viable System Model)
The prehistoric blueprint of the Cognitive Enterprise: 5 systems that map to multi-agent architecture.