The words, and how much they are worth.

Recall degrades as input grows, even when the task stays exactly as hard.

Summarising a full conversation and restarting. Selection is the hard part.

Load detail in tiers, so only the branch actually reached gets paid for.

A reference the agent holds, naming material it does not — plus when to go get it.

A pretrained concept the agent thinks with — frontier, fog, tracer bullet, lane.

Two budgets most writing collapses into one. Only the first should be minimised.

An instruction the model already obeys. It pays load to say nothing.

What tells an agent the work is finished. Vague bounds invite premature completion.

A README for agents. Plain markdown, no schema, closest thing to a neutral convention.

A folder of instructions an agent loads on demand. The portability is real.

A delegated worker in its own window, returning a summary rather than its transcript.

Open standard for connecting agents to external systems. The window cost is real.

A shell command the harness runs at a fixed point. Exit codes are the interface.

How much the agent may do without asking. A blast-radius control, not a correctness one.

Reuse of a processed prefix. Makes the front of a context expensive to change.

Agents compromised through the content they exist to process. No malware required.

Grep on demand or embed the repo. Genuinely unsettled — treat certainty as a tell.

Nodes and edges recording what blocks what. The useful part is the edge, not the diagram.

The nodes whose prerequisites are all satisfied — the edge of what is takeable now.

Work whose open questions gate each other. Density picks the instrument.

A thin slice all the way through, not a layer that goes nowhere alone.

The work item's body is the spec. Ready means a stranger could take it.

One session, one workspace, one branch, exactly one work item.

The vendor word for a lane, with the mechanics shipped. Guidance is three to five.

Record ownership before touching anything, in a form a machine can read.

Routes work, audits results, owns integration, never implements.

Separate directories over one object store. Not a security boundary.

One small file so a fresh session resumes with no re-explanation. Overwrite, never append.

A handoff points at the query and never enumerates items. Lists go stale on claim.

A check the model cannot argue past, because the harness enforces it.

An independent source of truth about correctness. Its presence predicts more than topology.

A claim that verification passed, offered instead of evidence. Re-run it.

A model grading a model. Standard, biased, and worst exactly when candidates are close.

Several reviewers with different lenses, rather than one working sequentially.

Agreement between models is not automatically evidence.

An explicit bucket for what a review leaned on but did not check.

The named person who adjudicates. Always asked, never inferred.

A 2026 name for task-graph orchestration. The patterns predate it by 19 months.

Designing the system that prompts the agent, instead of prompting it yourself.

Directing fallible agents under human oversight. Automates what you can verify.

Index the repo as symbols and relationships. Building it is the easy half.

Write the spec first and generate from it. Real tooling; oversold as executable.