This glossary translates current AI-engineering terms into decisions a product team can use. Each definition says what the term means in practice. The source label tells you whether the term comes from an inspectable artifact, research, an operator's account, or SWFT's own vocabulary.
Agent
An AI system that can choose actions and use tools while pursuing a goal inside defined limits. The model is one component of the agent, not the whole product.
How we know: First-party report (SELF-REPORT). Sources: Building effective agents.
Agent harness
The software around a model that supplies instructions, tools, context, memory, permissions, execution, and feedback. The harness determines what the agent can see, do, check, and recover from.
How we know: First-party report (SELF-REPORT). Sources: Harness Engineering; Demystifying evals for AI agents.
Agent productivity
Useful results produced by agents relative to delivery time, total cost, human correction, and quality. Tokens, turns, and generated code are activity, not productivity by themselves.
How we know: Analysis (INFERENCE). Sources: The rise of software factories; Running a Software Factory Efficiently at Uber Scale.
Agent skill
A versioned recipe with instructions, scripts, and reference material for a repeatable kind of work. The agent loads it when needed instead of carrying every procedure all the time.
How we know: First-party report (SELF-REPORT). Sources: Notion's Token Town; Vercel: agents as a new kind of software.
Agentic engineering
Designing the work, tools, context, permissions, and feedback that let AI agents complete useful goals. It treats the whole operating system around the model as a product.
How we know: Observed artifact (OBS). Sources: AI Engineer World's Fair 2026.
Attention gate
SWFT's name for a human approval gate: a point where a person must decide, such as approving a product choice, sensitive permission, merge, or production release.
How we know: Analysis (INFERENCE). Sources: Harness Engineering.
Autonomy
How far an agent can progress without help on a named kind of work at a measured success rate. Autonomy is not one permanent property of a model.
How we know: Independent study (STUDY). Sources: Measuring AI Ability to Complete Long Tasks.
Backpressure
A specific automatic rejection, such as a failing test or policy check, that returns bad work before it travels farther. Useful backpressure tells the agent what failed so it can correct the work.
How we know: First-party report (SELF-REPORT). Sources: Harness Engineering.
Coding agent
An agent that can inspect a codebase, make changes, run engineering tools and tests, and report what happened for a bounded software task.
How we know: First-party report (SELF-REPORT). Sources: Building effective agents.
Context engineering
Choosing what information an agent sees at each step, what it may retrieve later, and what should be removed or summarized so the next decision is well grounded.
How we know: Analysis (INFERENCE). Sources: Running a Software Factory Efficiently at Uber Scale; Building effective agents.
Context window
The amount of information a model can consider in one call. A larger window does not decide what is relevant; context engineering still has to select, retrieve, and summarize what the agent needs.
How we know: Analysis (INFERENCE). Sources: Running a Software Factory Efficiently at Uber Scale; Vercel: agents as a new kind of software.
Control plane
The part of a system that decides what may happen. It manages identity, permissions, budgets, queues, stop rules, and release policy even when an agent proposes the action.
How we know: Analysis (INFERENCE). SWFT editorial definition.
Eval-driven development
Improving an AI feature against a repeatable set of realistic scenarios and scoring rules. The scorecard is designed before broad rollout and gains new cases when failures appear.
How we know: First-party report (SELF-REPORT). Sources: Demystifying evals for AI agents; Notion's Token Town.
Evaluation
A repeatable assessment of an AI system on realistic tasks with explicit criteria and scoring rules. A related set of evaluations is an eval suite.
How we know: First-party report (SELF-REPORT). Sources: Demystifying evals for AI agents.
Evaluator
A program, model, or person that judges a result or run against explicit criteria. The evaluator itself should be checked against real human judgment.
How we know: First-party report (SELF-REPORT). Sources: Demystifying evals for AI agents.
Evidence package
The task brief, supplied context, run record, exact result, checks, approvals, release, and production observations supporting a claim that work is complete.
How we know: Analysis (INFERENCE). Sources: Demystifying evals for AI agents.
Factory throughput
The rate at which accepted software results reach their intended users or systems, measured beside quality and product value.
How we know: Analysis (INFERENCE). Sources: The rise of software factories; State of AI-assisted Software Development 2025.
Grader
One scoring component inside an eval, such as a test, a model following a rubric, or a person's judgment.
How we know: First-party report (SELF-REPORT). Sources: Demystifying evals for AI agents.
Guardrail
A deterministic limit on what an agent may do, spend, change, or release. Good guardrails include an explicit stop or escalation path.
How we know: Analysis (INFERENCE). SWFT editorial definition.
Harness engineering
Improving the system around the model, including repository structure, context, tools, permissions, feedback, and recovery, instead of only changing prompts or models.
How we know: First-party report (SELF-REPORT). Sources: Harness Engineering.
Human steering
The named moments when a person chooses a goal, resolves ambiguity, approves risk, corrects direction, interrupts a run, or owns a final decision.
How we know: Analysis (INFERENCE). Sources: Harness Engineering; Towards AI as a Collaborative Partner: A Taxonomy of AI Agent Behavior in Software Engineering.
Inner loop
The cycle that performs one job: make a change, check it, and correct it. A good inner loop gives the agent fast, trustworthy feedback.
How we know: Analysis (INFERENCE). Sources: AIEWF daily dispatch: loops and software factories.
Learning loop
The outer loop that studies repeated failures and outcomes, then improves future briefs, context, tools, evals, or rules. Fixing one result is recovery; improving the system is learning.
How we know: Analysis (INFERENCE). Sources: AIEWF daily dispatch: loops and software factories.
Managed agent
An agent operated as an owned service for one kind of work, with defined triggers, permissions, measurements, and results.
How we know: First-party report (SELF-REPORT). Sources: Running a Software Factory Efficiently at Uber Scale.
Model Context Protocol (MCP)
An open protocol that lets AI applications discover and use external tools and data sources. MCP provides connectivity; the application still has to choose relevant context and enforce permission.
How we know: Observed artifact (OBS). Sources: Model Context Protocol specification 2026-07-28.
Observability
The ability to inspect an agent's current state and run history: what it saw, which tools it called, what changed, what failed, how long it took, and what it cost.
How we know: First-party report (SELF-REPORT). Sources: Demystifying evals for AI agents; Harness Engineering.
Orchestration
Routing work among agents and deciding what happens when a worker completes, blocks, stalls, or fails. Good orchestration keeps task state and recovery rules outside any one agent session.
How we know: First-party report (SELF-REPORT). Sources: Building effective agents.
Proactivity
Useful work an agent begins from a schedule or a relevant situation instead of a direct request. Activity is only proactive when it is timely, grounded, and worth the attention it consumes.
How we know: Independent study (STUDY). Sources: Agentic Coding Needs Proactivity, Not Just Autonomy.
Production line
A repeatable path for one kind of job, from the task brief through checked delivery and feedback. A first factory pilot usually builds one line.
How we know: Analysis (INFERENCE). SWFT editorial definition.
Proof loop
SWFT's name for the agent making, checking, and correcting its own work before asking a person or a scarce shared system to review it.
How we know: Analysis (INFERENCE). Sources: Demystifying evals for AI agents; Notion's Token Town.
Run evidence
The trace, result, checks, decisions, and production observations kept for one agent run. Run evidence lets a reviewer verify what happened and lets a later attempt learn from failure.
How we know: Analysis (INFERENCE). Sources: Demystifying evals for AI agents.
Sandbox
An isolated workspace where an agent can run code with limited access to files, networks, credentials, and production systems. It reduces the damage a mistake can cause.
How we know: First-party report (SELF-REPORT). Sources: Vercel: agents as a new kind of software; AI Engineer World's Fair 2026.
Software factory
A repeatable way to turn a clear product request into working software. People choose the goal and quality bar; agents do bounded work; checks and human judgment decide what ships.
How we know: Analysis (INFERENCE). Sources: Robert W. Bemer; Running a Software Factory Efficiently at Uber Scale; Cursor's AI software factory.
Subagent
A bounded specialist agent given one part of a larger job. Subagents help when work is truly separable or parallel, but they add coordination, cost, and new failure points.
How we know: Analysis (INFERENCE). Sources: Building effective agents.
Task brief
The goal, allowed scope, settled decisions, open questions, owner, risk, required evidence, and definition of done for one job. It tells the agent what to achieve and what it must not guess.
How we know: Analysis (INFERENCE). SWFT editorial definition.
Task-completion time horizon
A reliability measure describing how long a task takes a human expert and the probability that an AI system completes comparable benchmark tasks successfully.
How we know: Independent study (STUDY). Sources: Measuring AI Ability to Complete Long Tasks.
Token efficiency
Useful accepted results relative to the model tokens consumed across the complete job. It is a diagnostic, not a replacement for quality, human attention, or product value.
How we know: Analysis (INFERENCE). Sources: The rise of software factories; Running a Software Factory Efficiently at Uber Scale.
Vibe coding
Generating software while deliberately paying little attention to the code itself, in Andrej Karpathy's original sense. It can help with disposable exploration but does not supply maintenance, proof, or release controls by itself.
How we know: Analysis (INFERENCE). Sources: Harness Engineering Is Not Enough: Why Software Factories Fail.
Work graph
A durable record of tasks, dependencies, blockers, ownership, and history outside any one agent session. It gives people and orchestrators a shared view of what can start next and what needs attention.
How we know: Analysis (INFERENCE). SWFT editorial definition.
Workflow
A model-and-tool system whose steps follow a path defined in ordinary code. Workflows are often preferable when the steps are known and predictability matters.
How we know: First-party report (SELF-REPORT). Sources: Building effective agents.
How do these definitions change?
SWFT updates a definition when the field changes, stronger evidence appears, or a plainer boundary becomes possible. Material changes receive a modified date and a correction note when the earlier wording could have misled a reader. A fashionable term may disappear; the underlying product decision should remain understandable.