Appendix - Agent Engineering

What We Learned

• The harness supplies execution, context, persistence and verification around a model.
• Choose loops or graphs from task structure, branching and recovery requirements.
• Retrieve context on demand; compact, offload and isolate information during long runs.
• Prepare coding and browser environments with clear instructions, permissions and fast checks.
• Use skills for reusable procedures, governed memory for facts and verifiers for completion.
• Write testable specifications; protect reward checks against gaming in agent training.

Key Acronyms, Concepts & Jargon

Term	Short meaning
Harness	The loop, tools, state, context and checks surrounding the model.
Loop / graph	Repeat actions until a condition holds / explicit steps and transitions.
Compaction / offloading	Compress context / move details to external storage.
Progressive disclosure	Loads a resource's details only when they are needed.
AGENTS.md / SKILL.md	Repository instructions / a reusable skill's instructions and trigger metadata.
Worktree	Separate working directory for a Git checkout, useful for isolated changes.
DOM / accessibility tree	Document Object Model / semantic UI representation used for automation.
Verifier / rubric	Checks task completion / explicit criteria for grading it.
Environment / reset	Task world, tools and feedback / return that world to a known starting state.
RLVR	Reinforcement Learning with Verifiable Rewards: train from checkable outcomes.
Reward hacking	Exploits a scoring mechanism without achieving the intended goal.
Acceptance criteria	Observable conditions that define whether a requirement is satisfied.
Spec-driven development	Builds from explicit requirements, plans, tasks and verification.
Stop hook	A check triggered before the harness permits a run to finish.

Back to section overview