Glossary

A

A2A (Agent2Agent)

An open protocol for agents to delegate tasks to other agents. Ogun plans to expose a workflow agent over A2A through a bridge process, off by default. It is not built.

ADE-bench

A benchmark by dbt Labs for AI agents doing analytics and data engineering work: tasks in dbt projects with a DuckDB or Snowflake database, modelled on Terminal-Bench. Ogun has an adapter; its task format is not yet verified.

Approval

A decision only a person can make, bound to a hash of the exact request. In Ogun: a new or changed code hash, a larger budget, a first model operator, or a node whose policy asks every time. Egress and publish approvals are reserved.

Artifact

An immutable materialization with an art_ id and a content hash, plus its source, parents, producing operation, run and verification state.

Asset

A logical object with versions over time, such as a weekly dataset. Each version points at artifacts.

B

Backfill

Running a partitioned workflow over a range of past partitions, for example every day since 2020, with per-partition status and resumption.

Benchmark profile

A run profile that switches off helpful behaviour (auto-repair, extra retries, caches, memory) and enforces isolation, so the measurement is of the agent, not of Ogun.

BLAKE3

A fast cryptographic hash. Ogun uses it for workflow hashes (over JCS) and alongside SHA-256 for artifacts.

C

Content-addressed store (CAS)

Storage where an object’s address is its hash, so identical content is stored once and any change is detectable. In Ogun it is the zeus-artifact store format of zeus-core, shared with Zeus: objects/blake3/ab/cd/<digest>.

Cost per success

Total cost of a set of runs divided by the number of successful tasks; a fairer comparison than pass rate alone.

Croissant

A metadata format for machine-learning datasets (MLCommons). The Zeus manifest exports can produce Croissant shape; Ogun does not export it yet.

D

DAG

Directed acyclic graph: nodes with dependencies and no cycles. A workflow is a DAG of operator nodes.

Determinism class

What an operator declares about repeatability: pure (same inputs, same outputs), seeded, or non-deterministic (network, models). It decides what can be cached and replayed.

E

Environment

The isolated place a run executes: base image, fixtures, workspace, tool policy. A first-class object, which is what makes benchmarks just runs.

Evaluator

Something that scores a run’s outputs, such as an official benchmark grader. Ogun records its version with the score.

H

Harbor

The harness and task format in which Terminal-Bench 2.0 is distributed. Ogun plans to read Harbor-format tasks through an adapter.

Harness gain

The average improvement a harness gives across models compared with a baseline harness: how much capability poor tooling was hiding.

I

Intermediate representation (IR)

The single JSON form every workflow compiles to, whether written in YAML, Python or natural language. It is what Ogun validates, hashes, diffs and runs.

J

JCS (RFC 8785)

The JSON Canonicalization Scheme: one byte-exact serialization of a JSON value, so hashing a workflow is stable. It is why the IR has no floating-point numbers.

L

Lineage

The graph of which artifacts and operations produced which outputs. Ogun records it for every artifact and can export it as OpenLineage events.

M

MCP (Model Context Protocol)

An open standard for giving AI applications tools and data. Ogun offers a small, hierarchical tool set so agents can create, run and inspect workflows.

O

OpenLineage

An open standard for lineage events (runs, jobs, datasets). Ogun can emit them so existing lineage tools understand its runs.

Operator

A unit of work in a workflow: acquire, extract, DuckDB, dbt, Python, OCR, an LLM call. Operators declare inputs, outputs, version, determinism and a sandbox profile.

Operator ABI

The contract between Ogun and an operator: a manifest plus an invocation protocol (a process speaking JSON over stdio first; WASI components are being considered).

P

Partition

A slice of a workflow’s work, by date or key. Partitions succeed, fail and retry independently, which makes large backfills practical.

Plan diff

What changes between two workflow versions: nodes, operators, sources, budgets and code hashes added or removed. It is what a person reviews before approving.

Provenance

Where data came from and how it was made: source (with secrets redacted), time, tools and versions, models and prompts, parents.

R

Run

One durable execution of a workflow: node instances, artifact edges, checkpoints, retries, events and provenance. A run survives crashes.

S

Sandbox

The restricted process an operator runs in: declared filesystem access, network policy and resource limits. Code never runs inside the daemon.

Session bridge

How Ogun reaches sources behind a browser login without holding cookies: it asks the Zeus daemon, as a service client, to acquire them.

Spider 2.0-DBT

The dbt subset of the Spider 2.0 benchmark: 68 tasks in which agents work on dbt projects backed by DuckDB.

T

Terminal-Bench

A benchmark of hard, realistic command-line tasks for agents; version 2.0 is distributed in the Harbor format.

Tool ablation

Removing one tool from a harness at a time and measuring the score drop, to learn which affordances actually help.

Trajectory

The replayable event stream of an agent run: model requests and responses, tool calls, commands, file reads and patches, database state hashes.

Trigger

What starts a run: time, an event, a data change, another agent, or a condition such as being on mains power.

W

W3C PROV

The W3C standard model for provenance (entities, activities, agents). Ogun’s records map onto it.

Workflow hash

BLAKE3 over the JCS form of a workflow. Approvals and runs are bound to it, so a changed workflow needs a new approval.

Z

zeus-core

The Rust engine library shared by Zeus and Ogun: transfers, resume, journal, verification, filesystem safety and the artifact store format.