AI Agents

An agent differs from a script in one respect that matters: it decides what to do next. AI Agents take a goal stated in ordinary language, break it into steps, choose tools to call, read what comes back, and revise as they go. The category holds general assistants, coding agents that edit repositories and run tests, research agents that reconcile sources, support agents that resolve tickets end to end, and frameworks for building bespoke ones. Tool calling, memory and guardrails are the recurring terms.

Engineering teams apply them to code review, migrations and on-call triage; support teams to ticket resolution; analysts to evidence gathering across systems. AI agent tools differ on which models they support, what tools and data sources connect, whether memory persists between sessions, how narrowly credentials are scoped, whether each run leaves a readable trace, and where a person must approve before something irreversible is sent, paid or deleted.

Judge candidates on a task whose correct answer you already know, then repeat it, because non-determinism means one good run proves little. Failure modes are specific: looping, fabricated tool arguments, confident wrong conclusions, runaway cost, and acting on instructions embedded in content the agent reads. Read-only access and staged approvals contain most of that. Pricing leans on usage, billed by tokens or actions, sometimes with per-seat licenses or a free allowance.