Agents

you stop scripting the steps and let the model choose which tool to call next, in a loop, until it decides it is finished.

Agent

the strongest agent architectures are mostly workflow, with agency confined to the steps that need it.

Workflow vs agent

thought, tool call, result, repeat, which is what your loop is already doing whether or not you call it by the paper's name.

ReAct pattern

the agent wrote out the steps first, then the third result contradicted the plan and it had to re-plan instead of carrying on.

Planning and decomposition

the sequencing, tool execution, retries, budgets and human checkpoints around the model calls are a few hundred lines of ordinary code you own.

Orchestration

the model remembers nothing, so everything the run has established lives in your application and gets re-presented every turn.

State management

the memory feature you shipped is a database plus a retrieval step, re-pasted into the window every turn.

Memory (short-term vs long-term)

you spend the afternoon deciding what to keep, summarise and throw out of the window each step, not rewording the prompt.

Context engineering

the agent was sharp for the first twenty steps and was making bad calls by step sixty, with the window full of dead ends.

Context rot

you replay the whole recorded run, thought by tool call by result, to find the step where it went wrong.

Trajectory

you cap the run at twenty iterations and a dollar, because a confused agent will otherwise loop until the invoice notices.

Stop conditions

the long run died forty minutes in and picked up from the last saved step instead of starting the whole thing again.

Checkpointing and resumability

you hand the sub-task to a child with a fresh context and get back only its conclusion, keeping the parent's window small.

Subagent

you split the work across five agents, spent five times the tokens, and one well-briefed agent did it better.

Multi-agent system

the second agent picked up the task without half of what the first one had learned, and redid the work or got it wrong.

Handoff

there is no API, so the agent takes a screenshot, clicks the button and types into the field, and breaks the day the layout changes.

Computer use / browser agents

the agent runs its shell commands inside a container against a scratch copy, so a bad step cannot touch production.

Sandboxing

you give the agent read-only tools by default and a scoped credential, because whatever its tools can reach is the blast radius.

Least privilege for tools

you put deterministic checks around the model, allowlists and schema validation and refusal filters, because you cannot rely on it policing itself.

Guardrails

you make the run stop and wait for someone to approve before it sends the email, writes to the database, or spends the money.

Human in the loop