Appendix A Glossary

Alphabetical order.

AGENTS.md / CLAUDE.md — a coding agent's project-level instruction file; "architecture documentation written for the agent." In practice it plays the immune system: one rule added per incident. ~100 lines as a table of contents is best practice.

Blackboard — a multi-agent pattern where all sub-agents write intermediate products to a shared workspace and a controller decides at runtime who acts; for tasks whose complexity grows over time with unclear role boundaries.

Build to Delete — design every harness component to be removable, with the system still working after removal; run scheduled deletion drills. Without a deletion mechanism, do not start building.

Closed loop — a loop with a verifier gate, stop conditions, rollback, and human intervention; opposite of an open loop (scheduled re-runs without feedback).

Compaction vs Reset — shorten history within one session (compaction) versus starting a fresh context in a new agent with a file-based handoff (reset). Some models only truly recover with reset.

Conditional edge — a graph edge whose target is decided at runtime by a (pure) routing function reading the state; the essence of ReAct.

Context Rot — systematic decision-quality decline once irrelevant content dominates the window; rule-of-thumb threshold ~40% fill.

CP checkpoint — compress a finished stage into a fixed-format summary (conclusions + key facts + open items) before handing to the next stage; never pass whole conversation histories downstream.

Ensemble — run redundant parallel copies of the same task and fuse by vote; trading redundancy for reliability.

Fail-closed — when the guardrail rule engine errors, deny by default rather than allow.

Fan-out / Fan-in — one node triggers several parallel nodes (fan-out); a merge node waits for all predecessors (fan-in).

feature_list.json — the core artifact of Anthropic's long-horizon pattern: a feature list with passes: false/true states. JSON over Markdown.

First Plausible Solution Bias — the model's attachment to its first passable plan; "looks OK" self-review as a stopping rule. Countermeasure: pre-completion checklists.

Factory Model — multiple loops cooperating to produce software: requirements/implementation/review/deployment loops, with deterministic gates passing artifacts between them.

Generator-Evaluator separation — the maker and the checker must be different instances; self-evaluation is systematically optimistic. Three evaluator forms: environment verification, deterministic tools, independent model.

Git checkpoint — incremental commits as state snapshots; the default rollback mechanism.

Hook — an automatic interception on agent lifecycle events (e.g., PostToolUse checks). "Hook > documentation": prose asks; hooks prevent.

Harness — the complete control infrastructure around a model: context supply, tools and permissions, state and memory, observability, human governance. Central formula: Agent = Model + Harness.

Harness decay — after model upgrades, old harness components turn from assets to friction (e.g., removing sprint decomposition after Opus 4.5→4.6 saved 38%).

JSONL inbox — a sub-agent communication protocol: one append-only JSONL inbox file per sub-agent; the orchestrator writes tasks; sub-agents poll.

LLM-as-a-Verifier — engineering the model-as-judge: finer score granularity, expectation over K repeated verifications, spec/output/error decomposition. Terminal-Bench 2.0: tie rate 27%→0%, accuracy 86.4%.

Loop Engineering — building systems that drive agents in autonomous loops rather than writing single prompts. Loop = trigger + decision + verification + budget + state. Hierarchy: Loop > Harness > Context > Prompt.

L1–L6 task ladder — autonomy levels by verifiability. L1/L2 fully automatic; L3 sampled; L4 plan approval; L5 per-action human confirmation (money/identity); L6 human-led.

Markov blanket injection — inject only context causally relevant to the current decision (contrast: injecting all 143 graph relations doubled false positives).

MCP (Model Context Protocol) — Anthropic's model↔tools/data connection protocol (late 2024). Representation cost ~32× a CLI; do the token math before wiring.

MVH (Minimum Viable Harness) — five starters: AGENTS.md, verification commands, feature_list.json, progress file, completion gate.

One-shot Syndrome — in long tasks, the agent sees partial progress after context exhaustion and declares completion without verifying.

Open loop — a feedback-free loop (e.g., pure cron re-runs). The agent repeats and self-confirms its errors until the budget dies; guaranteed failure.

Orchestration tax — parallel agent output outpacing human review bandwidth. Review bandwidth is the real cap on parallelism.

Outer loop — cross-session persistence of lessons: the right lessons, at the right granularity, written to the right place (AGENTS.md, progress files, rule libraries).

Pass@k / Pass^k — at least one of k succeeds (exploration ceiling) / all k succeed (release-grade consistency). Production decisions read the latter.

Blueprint — Stripe's pattern: deterministic nodes sandwiching probabilistic LLM work; at most N rounds (e.g., 2 CI attempts), then a human.

Progressive disclosure — inject names and descriptions by default; load full content on demand. Context engineering's first lever.

Ralph Loop — a 2025 structured loop-verification practice; precursor of the 2026 productized /goal and /loop.

Reducer — the per-field merge function in graph orchestration: nodes return deltas; reducers merge them into a new immutable snapshot.

Salt Signatures — Google's model-layer state route: encrypted reasoning state emitted before tool calls and restored later; parallel to harness-layer external artifacts.

Sandwich reasoning — high reasoning for planning + medium for implementation + high for verification (66.5%) beats uniformly high (53.9%).

Shadow validation — old and new systems in parallel with full comparison (one production system: 22,792 calls, 99.49% agreement). A transition instrument, not a permanent architecture.

Skills — artifacts that solidify project knowledge and operating patterns (SKILL.md); the carrier of procedural memory. The authoring format, versus Plugin (distribution) and Bundle (deployment).

Sovereignty — the five workstreams that do not vanish with model progress: will injection, permission granting, environment supply, boundary drawing, governance and audit. Sovereignty cannot self-generate.

Spec-file-driven — a multi-agent coordination mode: agents exchange structured files instead of talking; traceable, auditable, recoverable, decoupled, isolatable.

Sprint contract — a negotiated "what does done mean" between generator and evaluator before implementation starts.

State machine triad — State (data snapshot) / Node (acts and updates state) / Edge (decides the next node). The ontology of graph orchestration.

Three-way retrieval fusion — BM25 + vectors + graph fused with RRF; recovers exact keywords (paths, error codes, SHAs) that vector search dilutes.

Token paralysis — too many tools, wasted attention on choosing. Fix: deterministic per-task filtering (Stripe 500→15).

Working set — the minimal context actually needed for the current decision; the core of "total context → filter → working set → decision."

WIP=1 — one feature, one PR at a time; the safest default concurrency.

doom loop — the agent fine-tunes the same bad approach without converging. Four tourniquets: count-hint → rollback → budget breaker → error write-back.