A glossary of agentic delivery: terms every engineering leader should know

A lot of the vocabulary around AI-assisted software delivery is being invented in real time, which means the same word can mean slightly different things depending on who is using it. This glossary covers the terms that come up most often when teams start running humans and agents on the same delivery pipeline, defined plainly, without vendor spin.

Agentic delivery

The general practice of running AI coding agents as active contributors within a real software delivery process, not just as autocomplete inside an editor, but as entities that pick up tasks, produce PRs, and move through review like a contributor would. Distinct from "AI-assisted coding," which usually just means a human using an AI tool to help write their own code faster.

Agent loop

A failure pattern where an agent repeatedly attempts a similar approach to a task without converging on a working solution, often burning meaningful time and cost without visible progress. Distinguishing a genuine loop from a task that is just legitimately hard and taking a while is one of the harder open problems in this space; duration and cost alone are not reliable signals without comparing against how similar tasks normally behave.

Blocked state (agent context)

A status indicating an agent (or human) cannot currently progress a task, due to a missing dependency, an external service issue, a required decision from someone else, and so on. The useful version of this tracks not just that something is blocked, but since when and why, since "blocked for twenty minutes" and "blocked for two days" are entirely different problems that a simple yes/no flag cannot distinguish.

Cost-per-feature

A measure of the total resource cost, typically token spend plus human review time, of delivering a single feature end to end, as opposed to a per-token or per-API-call cost in isolation. Increasingly used as an alternative or complement to story points once agents are doing meaningful delivery work, because it captures actual cost variance that a size estimate does not.

Delivery risk detection

Automated flagging of patterns that historically precede a bad release or a stalled feature, merged code with incomplete QA close to a release date, a task blocked far longer than similar tasks typically are, an agent looping without converging. Distinct from generic project-management alerts (like an approaching due date) in that it is based on cross-referencing multiple live signals rather than a single calendar check.

Execution log

A recorded, timestamped trail of the actual steps, prompts, and decisions an agent (or a human using an AI assistant) took while working on a task. Useful for review, debugging, and institutional memory, since it captures the reasoning behind a solution, not just the final diff.

Human-agent capacity board

A delivery view that shows both human contributors and AI agents side by side, with the same status categories (available, active, blocked, at risk), rather than tracking agent work in a separate system from human work. The core idea behind treating agents as first-class delivery participants rather than a bolt-on tool.

Override log

A record of every time a lead or manager manually overrides a system-suggested action, an assignment, a risk dismissal, along with the stated reason. Useful both as an audit trail and as a feedback signal for improving suggestion quality over time, without automatically retraining on every override, which risks overfitting to one or two vocal decision-makers.

Org context layer

A searchable, synthesized record of an organization's accumulated engineering knowledge, past features, PRs, prompt sessions, and decisions, made queryable in natural language rather than requiring someone to know exactly where a specific decision was documented. Distinct from a code search tool in that it spans decisions and reasoning, not just the code itself.

Prompt-level context

The record of what was actually asked of an AI coding assistant during a task, as opposed to just the resulting code. Increasingly treated as institutional knowledge worth preserving and searching, since it often contains the reasoning and dead ends that never make it into a commit message or a ticket comment.

Redaction pipeline

An automated process for scrubbing sensitive content, credentials, customer data, personal information, from stored prompt or execution data before it is made searchable across an organization, typically combined with periodic human audit sampling since automated redaction reliably misses edge cases.

Rework rate

The proportion of completed work that requires a follow-up fix within a defined window after being marked done or merged. A more honest measure of delivery quality than raw completion counts, since it captures work that looked finished but was not.

Triage inbox

A holding area for inbound signals, error monitoring alerts, customer support tickets, that might warrant becoming a tracked feature, but require a human decision before they are promoted into the delivery pipeline. Prevents automatic noise (a single transient error spike, for instance) from flooding a backlog with features nobody actually wanted created.

This list will grow as the vocabulary around agentic delivery keeps evolving, which it will, quickly. If a term you have heard used differently than defined here, that is a genuine sign of how unsettled this space still is, not necessarily a sign either definition is wrong.

SignalsAI is built around these concepts directly: a shared human-agent capacity board, risk detection, an org context layer, and full execution logging.