Appendix D: Glossary of Core Concepts and Supporting Terms
Every entry below is the definition used from its introducing chapter onward. Chapter 24 is the argued version of this appendix; this is the fast lookup.
The six core concepts
Only six ideas get full named-concept treatment across this book, deliberately - see Chapter 24 for why the discipline stays at six.
Control/autonomy dial (introduced Chapter 1, resolved Chapter 26). How much structure you impose versus how much you delegate to the model - tuned to the task’s cost of being wrong, not maxed out in either direction. Lineage: a novel framing, though it names a tension (imposed process versus delegated judgment) that predates this book by decades in general software and management thinking. What’s new is sizing the dial specifically to cost-of-being-wrong rather than to team preference or model confidence.
State contract (Chapter 5). The explicit agreement about who writes which state channel and how concurrent writes merge. Lineage: a renamed application of prior art - CRDT merge semantics and Design-by-Contract thinking, applied specifically to LangGraph’s reducer model, where “no reducer” is itself a contract (LastValue) whether or not anyone wrote one down.
Routing boundary (Chapter 6). Control flow lives on edges, never buried in node side effects; the model proposes a route, the graph validates and disposes. Lineage: a renamed application of the general principle that untrusted input should never directly select a code path - applied here to a model’s own output, treated as untrusted by the same discipline as any other external input.
Checkpoint membrane (Chapter 10). The boundary where irreversible side effects begin - before it, retries are free; after it, every retry needs idempotency, an outbox record, or compensation. Lineage: explicitly builds on the saga pattern and the classic idempotency-boundary idea from distributed systems; the contribution is naming the boundary as a first-class design decision in an agent’s own graph topology, not inventing the underlying technique.
Context budget (Chapter 12). An explicit per-turn allocation of the context window across system, history, retrieved, and scratch - enforced on the hot path, not hoped for. Lineage: a novel naming for a problem that became acute specifically with long-context LLMs; the underlying instinct (bound a shared, scarce resource explicitly) is old, but “context budget” as a named, enforced allocation is this book’s own packaging of it.
Memory horizon (Chapter 13). The explicit, per-fact decision of how long a piece of information must outlive the thread that produced it. Lineage: a renamed application of cache-TTL and short-term/long-term memory thinking (cognitive science’s own vocabulary, borrowed by caching systems long before this book), applied here to the checkpointer-versus-store split.
Supporting terms
Local vocabulary, coined to make one chapter’s argument land, then used consistently without re-explanation. There are twenty. Chapter 24’s supporting-term index ties sixteen of them to the core concept each one operationalizes. The other four are defined here only: the containment ladder and authority resurrection from Chapter 23’s containment argument, and the two terms that arrive after Chapter 24. This table is the definitions alone, in introduction order.
| Term | Chapter | Definition |
|---|---|---|
| Abstraction ladder | 2 | Start at the highest rung that’s still correct (create_agent); descend to a raw primitive only when a concrete requirement forces it. |
| State-transition discipline | 3 | Every node reads state, returns a delta, and routes - and does nothing that isn’t one of those three. |
| Authority surface | 7 | The real-world actions a tool list grants the model; every tool widens it, a loose arg schema widens it invisibly. |
| Agent-loop seam | 8 | The defined hook points around the agent loop where cross-cutting concerns attach without corrupting node or tool bodies. |
| Resumption boundary | 9 | A durable run continues from the last completed superstep’s checkpoint, not from the start. |
| Approval gate | 11 | A deliberate suspension placed just before a membrane crossing, resuming only on an explicit, recorded human decision. |
| Background reflection | 14 | Forming and consolidating long-term memory off the response path, so remembering never adds latency to answering. |
| Coordination tax | 15 | The cost a multi-agent system pays regardless of whether the split helps - multiplied tokens, added latency, a fanned-out debugging surface. |
| Handoff payload | 16 | The context a handoff deliberately carries to the next agent - the sub-task and only the state the receiver needs, not the whole conversation. |
| The superstep barrier | 17 | Every node fanned out in a superstep must complete before the graph advances - what makes fan-in deterministic, and why one slow branch delays the batch. |
| Planning loop | 18 | The agent-maintained todo list that drives a harness’s own sub-task sequencing, as opposed to an externally fixed routing graph. |
| The live channel | 19 | The disconnectable, transport-level tap a client uses to watch a run - not the run itself, which stays durable regardless of who’s watching. |
| Trace fidelity | 20 | Whether a trace alone, without re-running code, can reconstruct exactly what happened in a run - a capability that must be designed in, not assumed. |
| Path coverage | 21 | A dataset’s job to exercise the distinct routes through a system’s topology, not just distinct input questions that land on the same route. |
| Schema straddle | 22 | The window during a rolling deploy when old and new code run concurrently against the same shared checkpoints. |
| Injection surface | 23 | The set of untrusted content sources that can reach the model as if they were instructions rather than data. |
| Containment ladder | 23 | Four ways to stop a run already doing damage, cheapest recovery first: refuse the call, revoke the authority, drain or cancel the run, kill the process. |
| Authority resurrection | 23 | A run regaining a permission that was withdrawn, because the withdrawal was stored where a checkpoint restore could reach it; kept outside graph state, a revocation survives the restore. |
| The durability tax | 25 | The operational cost a LangGraph project takes on once it relies on durable execution: a checkpointer and its database, state on every channel that the serializer can round-trip, and drain-and-deploy machinery for rollouts. |
| The judgment/execution split | 26 | Model capability shifts the optimal scaffolding for judgment; execution guarantees (durability, idempotency, approval gates) don’t move, because they were never compensating for model weakness. |
One additional local term - controllability wall (Chapter 1) - is deliberately excluded from this table. It names the point on the prototype-to-production path where added capability stops helping and reliability starts dominating, but it was scoped to Chapter 1’s own argument and never reused book-wide; including it here would misrepresent it as chapter-spanning vocabulary, which Chapter 1 itself never claimed for it.