Glossary
60 terms from “Designing Agent Workflows: From One-Off Conversations to Multi-Step Automation.” Hover the first occurrence in the lesson for its definition.
| Term | Definition | Source |
|---|---|---|
| workflow | An executable script that breaks a complex task into steps, delegates each step to a fresh agent, and owns the coordination and control flow itself. | Alex Op: Claude Code Workflows and Deterministic Orchestration |
| orchestration | The script deciding the order, parallelism, and branching of a task, instead of the agent deciding the control flow for itself. | ClaudeWorld: What Is a Workflow? Multi-Agent Orchestration Explained |
| single conversation | One agent finishing a task within a single context window, where all history and results live inside that one conversation. | Claude Code docs: Workflows |
| subagent | The independent agent instance a workflow step launches, with its own context window, that finishes one subtask and returns a result. | ClaudeWorld: What Is a Workflow? Multi-Agent Orchestration Explained |
| fan-out-reduce | The core of parallel decomposition: spread work across many parallel subagents (fan out), then merge every result into one output (reduce). | MindStudio: Five Claude Code Agentic Workflow Patterns |
| step | A workflow's atomic operation, either an agent call (needs reasoning) or a deterministic function (data transforms, math). | Mae Capozzi: Building a Multi-Agent Orchestrator |
| state | All the information a workflow currently holds: which step it's on, each step's result, and what it needs next, like the workflow's memory. | MachineLearningMastery: 5 Architectural Patterns for Persistent Memory and State in AI Agents |
| context | The information handed to a single agent call, selectively pulled from workflow state, containing only what that step needs. | Chrono Innovation: Architecture for Scalable Agentic AI Workflows |
| workflow state | All the information about the current task, passed between steps but not across sessions, and persistable to a database for recovery. | MindStudio: Workflow State vs. Session State |
| Session State | The conversation history and transient context the agent manages for itself, which the workflow doesn't need to track. | MindStudio: Workflow State vs. Session State |
| checkpoint | Saving workflow state to external storage after a key step, so a crash resumes from the latest checkpoint instead of starting over. | MachineLearningMastery: 5 Architectural Patterns for Persistent Memory and State in AI Agents |
| state machine | Modeling a workflow with discrete states (init, processing, completed) and explicit transition rules, persisting each transition. | MindStudio: Workflow State vs. Session State |
| sequential decomposition | A decomposition strategy that splits a task into steps that must run in order, each depending on the previous step's output. | OneUpTime: How to Create a Task Decomposition |
| parallel decomposition | A decomposition strategy that splits a task into independent subtasks that can run at the same time, with results merged at the end. | OneUpTime: How to Create a Task Decomposition |
| hybrid decomposition | A decomposition strategy where high-level phases run in order and each phase's internal steps run in parallel, combining both strengths. | OneUpTime: How to Create a Task Decomposition |
| dependency | The input/output relationship between steps; if step B needs step A's output, B depends on A and can't run until A finishes. | ACONIC paper: Systematic LLM Task Decomposition |
| chain-of-thought | A prompting technique that asks the LLM to show its intermediate reasoning while decomposing or reasoning, instead of jumping straight to an answer. | ApX Machine Learning: Task Decomposition Strategies for LLM Agents |
| zero-shot | Asking the LLM to complete a task with no examples, relying on its pretrained knowledge. | ApX Machine Learning: Task Decomposition Strategies for LLM Agents |
| transient error | A temporary error, usually from network fluctuation, resource contention, or brief overload, that may succeed on retry. | Augment Code: How Async AI Agent Workflows Survive Failure |
| permanent error | A persistent error, usually from misconfiguration, missing permissions, or a code bug, that won't succeed on retry. | Agents Arcade: Error Handling in Agentic Systems |
| exponential backoff | A retry strategy where each retry's delay grows exponentially (1s, 2s, 4s, 8s...), giving the service more time to recover. | Augment Code: How Async AI Agent Workflows Survive Failure |
| jitter | Adding a random perturbation to the retry delay so multiple clients don't retry at the exact same instant, spreading the load. | Augment Code: How Async AI Agent Workflows Survive Failure |
| circuit breaker | When the error rate crosses a threshold, it pauses calls to the failing service and fails fast, avoiding wasted resources and cascading failure. | Vasanthan: Handling Failures in Agent-Based Workflows |
| compensating action | An operation that undoes a completed step, so when a workflow fails partway through it rolls back earlier steps to keep data consistent. | AWS Marketplace: Agent Orchestration |
| Saga pattern | Defining a forward action and a compensating action for every step, and on failure running the compensations in reverse to achieve a distributed transaction. | AWS Marketplace: Agent Orchestration |
| idempotency | The property where running an operation N times has the same effect as running it once, producing no side effects on repeats. | Agents Arcade: Error Handling in Agentic Systems |
| branch | A workflow choosing a different execution path based on a condition, like an if-else or switch statement. | Alex Op: Claude Code Workflows and Deterministic Orchestration |
| loop | A workflow repeating the same operation, such as a for loop (over each element) or a while loop (until a condition holds). | Alex Op: Claude Code Workflows and Deterministic Orchestration |
| concurrency cap | Limiting how many parallel tasks run at once to avoid exhausting resources (too many open files, a saturated connection pool). | Augment Code: How Async AI Agent Workflows Survive Failure |
| input/output contract | Each step clearly defining what input it accepts and what output it returns, as precise as a function signature. | AWS Marketplace: Agent Orchestration |
| observability | A workflow's ability to answer questions like 'how far along am I, what errors have I hit, where are the performance bottlenecks'. | ClaudFlow: 7 Patterns for Claude Code Workflow Automation |
| progress tracking | Recording how many steps a workflow has finished, how many remain, and when it's expected to complete. | MachineLearningMastery: 5 Architectural Patterns for Persistent Memory and State in AI Agents |
| error log | A record of every error that occurred in the workflow, including step name, error type, error message, and timestamp. | Vasanthan: Handling Failures in Agent-Based Workflows |
| performance profiling | Measuring how long each step takes to identify a workflow's performance bottleneck. | ClaudFlow: 7 Patterns for Claude Code Workflow Automation |
| inverted control flow | A workflow's core innovation: instead of the agent deciding what to do next, the script decides which agent to call next. | Alex Op: Claude Code Workflows and Deterministic Orchestration |
| deterministic orchestration | A workflow's execution path is decided by the script's control flow (loops, branches), so the same input yields the same path. | Alex Op: Claude Code Workflows and Deterministic Orchestration |
| context window | The maximum amount of text an agent can see at once, usually hundreds of thousands to a million tokens; past that it must truncate or forget. | ClaudeWorld: What Is a Workflow? Multi-Agent Orchestration Explained |
| context explosion | When each step's result keeps piling into the context, so it grows without bound until it exceeds the limit or degrades performance. | Ranjan Kumar: Building Agents That Remember |
| resetting context | Each step clears the context and passes only what the current step needs, avoiding context explosion. | Ranjan Kumar: Building Agents That Remember |
| accumulating context | Each step's result is added to the context so later steps see all history, which suits a final summary step. | Ranjan Kumar: Building Agents That Remember |
| external storage | Persisting workflow state to a database, Redis, or the file system, rather than keeping it only in in-memory variables. | Appamass: State Management Patterns for Reliable AI Agent Workflows |
| Phase checkpoints | Saving a checkpoint after each major phase of the workflow completes, rather than after every small operation. | MachineLearningMastery: 5 Architectural Patterns for Persistent Memory and State in AI Agents |
| recovery logic | After a workflow resumes from a checkpoint, it skips the already-completed steps and continues from the point of interruption. | MachineLearningMastery: 5 Architectural Patterns for Persistent Memory and State in AI Agents |
| human approval | The workflow pausing at a key decision point to wait for a human to confirm before continuing, common before production deploys or data deletion. | Appamass: State Management Patterns for Reliable AI Agent Workflows |
| batch | Splitting a large set of tasks into small groups processed one group at a time, with parallelism inside each group, to control concurrency and resource use. | Augment Code: How Async AI Agent Workflows Survive Failure |
| rate limit | An API's cap on requests per unit time; exceeding it returns a 429 error, and you must wait or slow your request rate. | Augment Code: How Async AI Agent Workflows Survive Failure |
| resource contention | Multiple parallel tasks accessing a limited resource (a database connection, a file handle) at once, causing some to fail or time out. | Augment Code: How Async AI Agent Workflows Survive Failure |
| cascading failure | One service's failure causing the services that depend on it to fail too, with the failure spreading outward. | Vasanthan: Handling Failures in Agent-Based Workflows |
| fail fast | On an error you clearly can't recover from (a permanent error, an open circuit breaker), failing immediately without retrying, to save resources. | Agents Arcade: Error Handling in Agentic Systems |
| graceful degradation | When a non-critical step fails, the workflow skips it and continues instead of failing as a whole. | Augment Code: How Async AI Agent Workflows Survive Failure |
| dependency graph | A graph with nodes for steps and arrows for dependencies, showing the dependencies and execution order between steps clearly. | ACONIC paper: Systematic LLM Task Decomposition |
| DAG | The formal version of a dependency graph: nodes are steps, directed edges are dependencies, and 'acyclic' guarantees no deadlock. | ACONIC paper: Systematic LLM Task Decomposition |
| leaf | A node in the dependency graph that depends on nothing else, so it can run first. | Kinde: Multi-Agent Workflows for Complex Refactoring |
| topological sort | Determining the execution order of steps from their dependencies, guaranteeing each step runs only after its dependencies are done. | ACONIC paper: Systematic LLM Task Decomposition |
| Code refactoring pipeline | A full pipeline that analyzes code, generates a refactoring plan, refactors in dependency order, and verifies with tests. | Kinde: Multi-Agent Workflows for Complex Refactoring |
| Doc generation pipeline | An automated flow that extracts APIs from code, generates examples, renders docs, and publishes them. | ClaudFlow: 7 Patterns for Claude Code Workflow Automation |
| Test automation flow | An end-to-end testing workflow that sets up environments, runs tests in parallel, collects results, and generates a report. | ClaudFlow: 7 Patterns for Claude Code Workflow Automation |
| task decomposition | Taking a vague, oversized task and breaking it into small, clear steps, each with defined inputs and outputs. | ApX Machine Learning: Task Decomposition Strategies for LLM Agents |
| Debugging a workflow | Locating problems in a workflow through logs, state inspection, and step-by-step tracing. | ClaudFlow: 7 Patterns for Claude Code Workflow Automation |
| half-open | The circuit breaker state where, after a timeout, a few trial requests go through: if they succeed the breaker closes, otherwise it stays open. | Vasanthan: Handling Failures in Agent-Based Workflows |