Reference

Last updated on 2026-09-16 | Edit this page

Glossary

Agentic coding
Working with an AI tool that can read a repository, execute commands, edit files, and iterate on its own output semi-autonomously — as opposed to suggest-and-accept assistants.
Agent loop
The cycle a harness runs: request → understand → plan → act → observe → revise → repeat. Productive when the task is well specified; a token-burning spiral when it isn’t.
Context window
The bounded working memory of a model. Everything — your prompt, file contents, command output, conversation history — competes for space and attention within it.
Context accumulation
The growth of conversation context over a session; each new message resends the accumulated history, so late messages cost far more than early ones.
Context file / project context
A file (CLAUDE.md, .github/copilot-instructions.md, AGENTS.md) loaded at session start that gives the agent persistent knowledge about a project: structure, conventions, commands, safety rules. Advisory, not enforced.
Data poisoning
Inserting crafted examples into a model’s training data to implant hidden behaviors. Research suggests a few hundred poisoned documents can suffice, roughly independent of model size.
Data leakage
Any path by which information from outside the training data (typically from the test set or the future) influences model training or evaluation. Produces inflated scores from code that runs cleanly.
Feature-based development
Delegating work to an agent one well-scoped, verifiable feature at a time — one feature per session, a small diff, one pull request — rather than project-sized requests. A feature is one thing you can check.
Group leakage
Data leakage caused by splitting on rows when the unit that repeats (subject, image, document, site) appears on both sides of the split. The model learns the unit, not the signal. Fix: split by group (GroupShuffleSplit, GroupKFold) and assert disjointness.
Harness
The application around a language model that lets it act on a real codebase: file access, terminal and test execution, context management, planning, tool integrations, memory, permission controls, and the agent loop. Agent = LLM + harness + tools + loop.
Hook
A deterministic script configured to run at a fixed point in an agent’s workflow (after every edit, before every commit). Unlike a context file or a skill, a hook is guaranteed to fire rather than advisory.
Model backdoor (trojaned model)
A model engineered — via poisoned training data, fine-tuning, or surgical weight edits — to behave normally except on a trigger input (e.g., emitting vulnerable code when a condition is met). Lives in the parameters, so file-format defenses like safetensors don’t help and no scanner or benchmark reliably detects it.
Instructions vs. permissions
An instruction (a rules file, a prompt) asks the model to behave and is text it weighs; a permission (a VM boundary, a deny rule, branch protection) removes the ability. Instructions shape behavior; permissions constrain it.
MCP (Model Context Protocol)
An open standard for plugging tools and data into any agent. Unlike an API, where you write the calling code, the agent discovers an MCP server’s tools and decides when to call them.
MCP server
A lightweight program that exposes tools, resources, and prompts to an agent over MCP, locally over stdio or remotely over HTTP. Its output enters the agent’s context as untrusted input.
Minimum viable pipeline (MVP)
Whatever you can get running quickly and understand end to end: a slice of the data, one model, your laptop. Fewer friction and failure points, and the ideal first plan for an agent.
Plan file (plan.md / PLANS.md)
A committed, living document holding goal, scope and acceptance criteria, ordered steps, progress, discoveries and blockers, and test results. The agent’s direction, your review point before implementation, and a shared definition of done.
Plan mode / read-only mode
An agent mode in which files can be read and questions answered but nothing is modified. The safest first contact with any repository.
Prompt injection
An attack in which content the agent reads (a README, issue, data file) contains instructions crafted to hijack the agent’s behavior.
Safetensors
A weights-only serialization format for model files that, unlike pickle-based formats, cannot execute code when loaded. Prefer it when downloading models.
Slopsquatting
Registering a malicious package under a name that LLMs habitually hallucinate, so that agents (or people following AI suggestions) install it. Cousin of typosquatting (lookalike names) and dependency confusion (shadowing an internal package name on a public registry).
Supply-chain attack
Compromising software by attacking something it depends on — a package, a build pipeline, a model file, or the agent tooling itself — rather than the software directly.
Skill
A reusable, on-demand instruction package for an agent (e.g., a deployment checklist or a lab’s analysis conventions) that loads only when relevant.
Test-driven development (TDD)
Writing a test that states the required behavior before writing the code, then writing (or having the agent write) code that makes it pass. With an agent, the test is both the specification and an executable feedback loop. For research code, the tests written first concern the data and the result rather than function signatures.
Test as contract
Encoding a requirement — especially a data property — as an executable test the agent must satisfy, turning specification into verification.
Underspecification
Leaving decisions unstated in a request. The agent resolves each gap with the most statistically typical (“average case”) choice, which is rarely your case.

Tool equivalents quick reference

Task Claude Code GitHub Copilot
Cloud VM session (nothing local) claude.ai/code Copilot app in cloud mode; github.com/copilot/agents; assign an issue to Copilot
Read-only exploration / plan review Plan mode (Shift+Tab; built in on the web) Explicit plan mode in the Copilot app; on the web, ask for a plan in the prompt; Ask mode in VS Code chat
Network allowlist Cloud environment Network access: None / Trusted / Full / Custom (own domain list) Repo Settings → Copilot → Cloud agent → Internet access (firewall, recommended/custom allowlist); repo admin only
Reset context between tasks /clear New chat
Tame a long conversation /compact Carry a summary into a fresh chat
See usage/cost /cost (and /context) Copilot settings usage
Switch model /model Model picker in the chat panel
Interrupt the agent Esc Stop button
Project context file CLAUDE.md (also reads AGENTS.md) .github/copilot-instructions.md (also reads AGENTS.md)
Command allow/deny rules /permissions, settings.json Terminal auto-approve settings
Add an MCP server claude mcp add …; claude mcp list MCP: Open User Configuration
Skills .claude/skills/<name>/SKILL.md; plugins via claude plugin install Prompt files .github/prompts/<name>.prompt.md

Further reading

Tools and workflow

Safety and security

Research