A software system that decides and acts on its own — using tools to read files, run commands, query systems, and edit code — rather than only answering when prompted. The word comes from Latin agens, "doing" or "driving"; the adjective form, agentic, names the shift this category represents: from generative AI that suggests work to AI that performs it. The mechanics are deceptively simple: a language model sees the conversation so far, decides which tool to call next, receives the output, and loops — perceive, decide, act, observe, repeat — until it judges the task done. The intelligence is the same model that powers a chatbot; what differs is the apparatus around it — tools, persistent context, a termination check. A droid is an agent specialised for one repeating task on a schedule — the same machinery, tuned for reliability over range.
A tangible output left behind by a process — a compiled binary, a build log, a generated image, an AI session transcript. The word is borrowed from archaeology, where it distinguishes objects made by human work from natural phenomena. In agentic coding, artifacts are what you can inspect after the fact: CI builds produce test reports as artifacts; AI sessions produce code diffs and conversation logs; an autonomous droid run produces a PR plus a journal entry. The distinction matters because an artifact can be reviewed and audited long after the producing process is gone.
The moment a deliberately fragile guardrail trips, warning operators that a forbidden pattern has recurred. Borrowed from coal miners, who carried caged canaries into shafts as living gas detectors — the bird dies first, the miners get out alive. In software, a canary is a small test or signal designed to fail loudly the instant something it watches degrades; saying the canary "fired" is shorthand for the warning having worked. Teams set canaries on flaky deployments, regression-prone code paths, and any rule they expect to be tested over time.
A traffic-jam failure that happens when several automated pull requests merge in quick succession, each prepared against the codebase as it looked before the others landed. The first merge goes through cleanly; everything behind it now conflicts on whatever file they all touch — often a shared changelog or history file that every job appends to. The fix is rarely more speed: it is sequencing, retries, or giving each writer its own file. Anywhere several automated jobs share one repository — changelog writers, dependency updaters, release-note generators — the cascade race is waiting.
A program that runs quietly in the background, waiting to act — handling work when some condition is met rather than when a person asks for it. The name comes from Maxwell's demon, a thought-experiment creature that sorts molecules tirelessly; early MIT programmers borrowed it for the helper processes that keep a system running unattended (it is pronounced "demon", and is sometimes backronymed "Disk And Execution Monitor"). A web server, a job scheduler, and a print spooler are all daemons: started once, they persist, listening for the event that gives them something to do. Long-running by design, a daemon is judged on reliability over weeks of uptime rather than on any single response — which is also why it makes a natural home for an autonomous agent that needs to run unattended.
A small, focused autonomous AI agent built to run a narrow task on a schedule rather than respond conversationally. The word is borrowed from Star Wars — itself a shortening of "android" — where droids are mission-specific helpers. In agentic coding, a droid differs from a general-purpose assistant in scope: one droid might watch only for security regressions, another only check accessibility, another only patch outdated dependencies. The constraints that distinguish a droid from an unconstrained agent — a tightly scoped prompt, a small per-PR change limit, a fixed nightly cadence — are what make the pattern safe to deploy on a production codebase.
The software framework that wraps an AI model and turns its decisions into practical work — reading files, running commands, holding context, routing tool calls — plus the parts that make the work reliable: permissions, verification, and recovery when a step fails. Borrowed from the gear that attaches a working animal to a load: the harness doesn't generate the power, it makes the power useful. The model produces text and decides what to do next; the harness decides what that's allowed to touch, and whether the result is checked before it counts. Claude Code, Cursor, and Replit Agent are different harnesses for similar models — and as models converge, the harness is increasingly where the practical advantage lives.
A design pattern where automation does the work but a human reviews, approves, or corrects the output before it ships. The term originated in machine learning, where "humans in the loop" labelled training data and validated model predictions. In agentic coding, HITL is what makes autonomous agents safe to deploy on a production codebase: the agent writes the change, the human decides whether it merges. The opposite extreme — full autonomy — trades safety for speed. Most professional agentic workflows today run HITL by default; full autonomy is a graduation, not a starting point.
No-op is short for "no operation" — an action that runs but deliberately changes nothing. The term comes from assembly language, where the NOP instruction occupies the processor for a cycle without altering any state. In agentic work a no-op is an agent that ran, checked, and correctly decided there was nothing to do — a recorded run that touches no code. That "nothing to do" outcome is a healthy signal, not a failure: it means the checks passed and no change was warranted.
A fixed set of scoring criteria for grading a model or agent's output against a defined standard — turning "is this any good?" into specific, repeatable checks. Each criterion is applied the same way across many outputs, by a human reviewer or a second model acting as judge, so quality becomes comparable rather than anecdotal. A rubric is what separates "the new version feels smarter" from "it met nine of ten criteria, up from six."
A minimal end-to-end check that a system is alive and the basics work — not deep validation, just a quick "does it light up?" confirmation. The term comes from hardware engineering: when an electrician plugs in a new circuit, if it doesn't catch fire, the wiring is at least connected correctly. In software, smoke tests run first because they fail fastest. A typical use is verifying a deploy: does the homepage load, does the database respond, does auth still work? If yes, run the deeper tests; if no, roll back before anyone notices.
The process of reviewing accumulated work — typically open pull requests or issues — to decide what to merge, close, or defer. Borrowed from emergency medicine, where it means rapidly sorting cases by urgency and viability. In agentic coding, triage is the human gate that decides which autonomous-agent output ships, which gets closed, and which signals a pattern worth fixing upstream (a prompt, a CI rule, or a cooldown). At Epic Growth we triage Jules droid PRs after every nightly run.
A mode of agentic coding, from Claude Code, where the AI orchestrates many subagents at once instead of working alone — fanning a task across parallel workers and adversarially checking their findings before committing. It trades cost for coverage and confidence, so it is reserved for large or high-stakes work: big migrations, audits, deep reviews where one context can't hold the whole problem.
In software, a sequence of automated steps that runs in response to a trigger — a code push, a schedule, a pull request, a manual button. The term is most visible in CI/CD systems like GitHub Actions, where workflows are written as YAML files and executed by hosted runners. A workflow can build code, run tests, deploy to a server, or coordinate autonomous agents. Whether the trigger fires once or every night, the workflow's value is that it replaces a human remembering to do the work.