Glossary

Adversarial review gives independent reviewers opposing roles and requires evidence for each objection.

slides49 66 67 seeverification fan-out review decay

An agent is a model loop that can use tools and check its work.

AI Coding Dictionary · exact match ↗ · close match in rune ↗

slides5 6 12 14 15 20 21 26 27 33 39 41 42 44 46 47 51 52 55 56 57 59 61 70 71 72 74 76 seeagentic loop harness

Agent engineering designs the loop, its tools, and its stop condition.

slides21 seeagentic loop harness engineering prompt engineering

An agent team has sessions that share one task list and send messages to each other.

slides14 57 76 seefan-out working copy

An agentic loop returns each model reply as input until a stop condition ends the run.

slides5 6 20 21 25 27 28 29 55 seeagent stop condition tool call

An app is a software artifact that an agent loop produces or operates.

slides7 15 27 seeagent artifact computer use

An artifact is a file on disk that an agent loop uses or produces.

close match in rune ↗

slides27 46 66 74 seeapp

Attention lets each token weigh itself and every earlier token, then combine their vectors.

slides17 18 22 30 seecontext window layer vector

Authenticity theater hides signs of agent work without improving the artifact or its verification.

slides46 seetheater test verification

Automatic execution runs a check without a person starting it, for example in a pre-commit hook or in CI. A deterministic check often runs this way, but it does not have to.

slides41 52 seeci parity hook pre-commit hook

A base model is the pre-trained model that continues the text it receives.

slides23 seeinstruction-tuned model post-training pre-training

A boundary stack combines controls that answer different questions about tools, system calls, network traffic, and custom policy.

slides32 40 70 seeegress cage hook permission sandbox

A cage is the containment stack that enforces permissions, sandbox rules, egress rules, and project access.

slides40 69 70 71 73 seeboundary stack credential broker telemetry

CI parity applies the same CI checks and standards to human and agent changes.

slides52 seeguard verification

Compaction replaces conversation history with a model-written summary to free context space.

AI Coding Dictionary · exact match ↗

slides4 28 29 30 seecontext management context window

A completion is the text or tool call that the model returns for one request.

slides19 seerequest tool call

Computer use lets an agent operate a real browser or a controlled desktop.

The agent changes the mass in the spacetime app and reports the clock rates.

slides15 seeagent app verification

Context is the information the agent has for its current task.

AI Coding Dictionary · exact match ↗

slides13 seeagent context management context window

Context management loads, compacts, and clears material in the model request.

slides20 29 seecompaction context window harness

A context window is the fixed-size text window that contains everything the model can read for one request.

AI Coding Dictionary · exact match ↗

slides4 17 18 30 60 seeattention compaction request

A credential broker holds a service token outside the cage, applies a read and write policy, and makes the call for the agent.

slides71 73 seecage permission

A deterministic check gives the same result for the same input.

AI Coding Dictionary · close match ↗

slides65 seeverification automatic execution judged rubric

A done condition is an independent check that shows whether the work succeeded.

slides25 seestop condition verification

Effort sets how much the model reasons before it answers. More effort spends more output tokens.

AI Coding Dictionary · exact match ↗

slides62 seemodel model routing token

An egress cage controls where bytes can leave and denies destinations that are not listed.

The cage blocks example.com and permits an allowlisted package host.

slides35 69 70 seeboundary stack cage network allowlist

Factuality is agreement with verifiable facts. A factuality failure is a false claim about the world.

slides31 seehallucination verification

Faithfulness is agreement with the source the model read. A faithfulness failure misstates that source.

slides31 seehallucination verification

Fan-out sends one question to several independent agents and combines their answers.

slides14 51 67 seeadversarial review agent team

A feed-forward step works on each token vector without reading other tokens.

slides22 seeattention layer

A fork starts a subagent with a copy of the current conversation, so it can reuse the prompt cache.

slides60 seesubagent prompt cache

A guard is hook code that blocks a dangerous command before execution.

The guard blocks rm -rf before the command runs.

slides36 38 40 54 72 seehook layer order

A hallucination is fluent, confident output that is false about the world, or that drifts from the source the model was given.

AI Coding Dictionary · exact match ↗

slides16 seefactuality faithfulness

A harness is everything around a model that makes it an agent: deterministic code that manages context, tools, permissions, and state, and the system prompt it sends.

AI Coding Dictionary · exact match ↗ · close match in rune ↗

slides13 16 19 20 21 26 27 28 39 41 44 48 52 55 63 64 seeagentic loop context management model permission system prompt

Harness engineering designs permissions, context, data, and agents around the loop.

slides21 seeagent engineering harness

The head turns the final vector into one score for each vocabulary entry.

slides22 seesoftmax token table vector

A hook is deterministic user code that runs at a seam in the agent loop.

close match in rune ↗

slides29 31 36 37 38 39 56 76 seeguard harness

An instruction file is a file of guidance that a harness loads for a user, a repository, or a directory.

close match in rune ↗

slides27 39 41 61 68 76 seeartifact harness skill

An instruction-tuned model is a post-trained model adjusted toward instructions and preferred replies.

slides23 seebase model post-training

Iteration repeats generation, factual checks, contradiction, and new generation.

slides11 74 seerecombination verification

A jj workspace is a separate working copy that jj manages, one for each agent.

slides57 58 seeworking copy

A judged rubric evaluates work that deterministic checks do not cover and marks each judgment for a person.

slides65 seeverification deterministic check

A model layer applies attention and then a feed-forward step to token vectors.

slides22 23 seeattention feed-forward step transformer

Layer order puts prose above enforcing hooks and deny rules. A higher layer never weakens a lower layer.

slides38 seeboundary stack guard

Managed settings are root-owned controls that the agent can read but cannot edit.

slides72 seecage self-modification wall

A matcher selects the tool and pattern that a permission or hook rule covers.

slides33 38 39 seehook permission

MCP is the Model Context Protocol, which connects a harness to external tool servers.

AI Coding Dictionary · exact match ↗

slides28 seetool call tool catalog

A model is a fixed set of trained numbers that predicts the next token.

AI Coding Dictionary · exact match ↗ · close match in rune ↗

slides3 4 5 6 9 11 13 16 17 18 19 20 22 23 24 25 26 28 30 31 40 42 47 49 54 61 63 64 68 71 seeharness request token training

Model character is a model's consistent response style, such as explanatory or defensive behavior.

slides9 63 64 68 seemodel model routing

Model routing assigns planning, execution, or review to a model selected for that task.

slides9 63 64 seeharness model model character

A network allowlist names the destinations that can receive bytes.

slides35 39 69 70 72 75 seeegress cage

A permission rule allows, asks about, or denies a tool call.

Ask before Bash runs git push.

AI Coding Dictionary · close match ↗

slides13 20 21 26 28 29 33 39 42 68 70 76 seematcher permission request sandbox

A permission request asks a person before the harness runs a selected tool call.

AI Coding Dictionary · exact match ↗

slides28 seepermission tool call

Post-training adjusts the same model numbers toward instructions and preferred replies.

People or a judge prefer the answer that uses ls -la.

slides23 24 31 seetraining instruction-tuned model sycophancy

A pre-commit hook runs project checks before a commit completes.

slides41 52 seeci parity deterministic check hook

Pre-training adjusts model numbers by predicting the next token across a text corpus.

The model predicts cell after The mitochondria is the powerhouse of the.

slides2 22 23 24 seetraining base model

A prompt is what a person or harness says to one model call.

slides8 9 10 13 15 20 21 24 28 29 32 38 42 47 68 76 seeprompt engineering request

A prompt cache lets a provider reuse the start of a request that it has already processed, at a lower price.

AI Coding Dictionary · exact match ↗

slides60 63 seeprovider request system prompt token

Prompt engineering controls what is said to one model call.

slides21 seeagent engineering prompt

Prompt injection is untrusted content that tries to direct the agent.

slides32 seeboundary stack permission

A provenance record states the origin of each deployed skill file.

close match in rune ↗

slides61 seeartifact skill

A provider serves a model for inference through a local or remote endpoint.

AI Coding Dictionary · exact match ↗

slides26 39 seeharness model

QAT trains the model while the number rounding is in place.

slides23 seequantization training

Quantization rounds model numbers to fewer bits and reduces the model file size.

The file uses about four bits for 9.3 billion numbers and about six bits for 2.6 billion.

slides23 seeqat training

Recombination joins patterns that already exist in the training data.

slides11 69 seeiteration training

A request is what the harness sends to the model: the latest input and the conversation history. The model's answer is a separate completion.

AI Coding Dictionary · close match ↗

slides3 8 9 13 19 28 49 76 seecompletion context window model prompt

A review bot is a second model that reviews a change, or one narrow part of it such as security, before a person does.

AI Coding Dictionary · close match ↗

slides40 seecage verification

Review decay is the loss of useful findings across repeated approval rounds.

slides66 seeadversarial review verification

A sandbox is an isolated environment that limits what can run and where it can run. It can be an operating-system boundary, a container, or a virtual machine.

The kernel rejects a write outside the allowed paths with EPERM.

AI Coding Dictionary · exact match ↗

slides28 34 39 68 70 73 seeboundary stack cage permission virtual machine

The self-modification wall prevents the agent from editing the settings that control it.

slides72 seeboundary stack cage managed settings

A session is one bounded run of work with an agent. It keeps the conversation across turns. Compaction shortens the history, and the session continues.

AI Coding Dictionary · close match ↗

slides13 seeagent compaction context window turn

A skill is a teachable capability that stays outside context until the agent needs it.

AI Coding Dictionary · exact match ↗ · close match in rune ↗

slides29 49 61 seeagent harness instruction file

The small change loop explores the code, plans a route, makes a small diff, and runs the project's tests.

slides42 seeagentic loop verification

Softmax converts the head's token scores into probabilities.

slides22 seehead token

A stop condition ends the agentic loop when the model stops calling tools.

slides6 20 21 25 seeagentic loop done condition

A subagent is an agent that another agent starts in its own context. It can start subagents of its own. Only its result returns to the parent.

AI Coding Dictionary · close match ↗

slides27 28 29 60 63 seeagent agentic loop

A surface is the terminal or editor window around a harness.

slides27 seeapp harness

Sycophancy is agreement encouraged by post-training rewards for agreeable answers.

AI Coding Dictionary · exact match ↗

slides31 seeadversarial review post-training

A system prompt is the instructions that the harness puts at the start of every request.

AI Coding Dictionary · exact match ↗

slides19 28 29 seeinstruction file request

Telemetry records sessions and commands but does not enforce limits.

slides40 70 seecage verification

The theater test asks whether the model can argue with a claimed limit.

slides40 71 seecage verification

A token is one text piece with an identifier that selects a vector from the token table.

The two occurrences of the token the both have ID 506.

AI Coding Dictionary · exact match ↗

slides16 17 18 22 23 30 54 60 seeattention token table vector

The token table maps each token identifier to its starting vector.

The table has 262,144 rows with 3,840 numbers in each row.

slides17 22 23 seehead token vector

A tool call is structured model output that asks the harness to run a named tool with specified arguments.

run_command with a JSON string that contains the command

AI Coding Dictionary · exact match ↗

slides6 19 26 28 30 seepermission request tool result tool schema

A tool catalog is the set of tools that the harness offers to the model.

slides29 seeharness skill tool schema

A tool result is the tool output that the harness appends to the conversation.

mean period: 2.0036 s

AI Coding Dictionary · exact match ↗

slides6 13 26 28 seeagentic loop tool call

A tool schema describes a callable tool and the structured input that it accepts.

slides19 26 seetool call tool catalog

Training sets a model's numbers from examples: pre-training for next-token prediction, then post-training for instructions and preferred answers.

AI Coding Dictionary · close match ↗

slides3 11 22 23 24 31 seemodel quantization

A transformer is a stack of repeated attention and feed-forward layers.

slides2 22 seeattention layer

A turn starts with a user message and ends when the agent hands control back. It can hold several requests.

AI Coding Dictionary · exact match ↗

slides3 19 63 seeagentic loop request

A vector is one row of numbers that represents a token at one processing stage.

slides17 18 22 seeattention head token table

Verification uses tests, reviews, sources, or probes that can prove generated work wrong.

slides5 31 42 43 45 46 47 48 50 61 66 69 73 75 seedone condition

Vibecoding accepts generated code that nobody read.

AI Coding Dictionary · exact match ↗

slides43 seeiteration verification

A virtual machine runs a separate operating system on emulated hardware. A program inside it cannot reach the host directly.

seecage

A working copy is one separate project directory assigned to one agent.

slides57 seeagent team artifact

A Git worktree is a separate working copy that shares one repository with other worktrees.

slides57 seeworking copy jj workspace