Coding Horizon

OMG: The Missing Ledger For AI Agent Teams Now

Every property and claim the finished video puts on screen, chased to a primary source. Where a shot draws an example rather than a measurement, that is stated here too, because an illustration that reads as a measurement is the same mistake as an unsourced number.

The project

OMG, published as “Oh My Group”. Repository jeremy-merchant/OMG, written in Go, and described by its own README as “Local first coordination and recovery for people and AI coding agents.” The executable is omg.

• Source: https://github.com/jeremy-merchant/OMG

What the ledger stores. The README states it “keeps lineage, tasks, runs, progress, dependencies, typed messages, handoffs, advisory path reservations, and observed Git state in one local SQLite ledger”. The video’s record table draws those groups with an example against each.

• Source: https://github.com/jeremy-merchant/OMG (README, “What the OMG CLI tracks”)

Local first, and daemonless by default. The README’s “How it works” section states that canonical coordination state lives in an owner-scoped SQLite store outside Git, and that no background service, cloud account, Node runtime, Python runtime or shared SQLite library is required. Both are drawn as the trust story in chapter 4.

• Source: https://github.com/jeremy-merchant/OMG (README, “How it works”)

Path reservations are advisory. The README describes “advisory path reservations” and states that they, together with dependency state, “make conflicts visible”. The video draws a reservation as three fields — path, worker, reason — which is what the README’s model amounts to; it is not a screenshot of the tool’s own output.

• Source: https://github.com/jeremy-merchant/OMG (README, “Why Oh My Group”)

Claimed done is kept separate from verified done. The README states the CLI distinguishes self-reported completion from independent verification and does not “Collapse WORK_COMPLETE and VERIFIED_DONE into one state”. This is the whole of chapter 2.

• Source: https://github.com/jeremy-merchant/OMG (README, “Safety boundary”)

The authority it does not take. The README’s safety table states the CLI does not “Grant commit, push, deploy, credential, production, deletion, or publication authority”, and does not “Commit, merge, rebase, push, reset, clean, delete, or remove branches/worktrees”. The video’s deny panels and both boundary shots draw exactly this list.

• Source: https://github.com/jeremy-merchant/OMG (README, “Safety boundary”)

Git is observed, not driven. The README states Git remains authoritative for code, commits, refs, branches, worktrees, diffs and history, and that the CLI stores coordination and policy facts plus explicitly requested evidence. Chapter 8 draws a read-only git status and the stale-state consequences that follow from watching rather than driving.

• Source: https://github.com/jeremy-merchant/OMG (README, “How it works”)

What is drawn as an example, and is not a measurement

These appear on screen as illustrations of the argument. They are invented for the video and describe no real repository, product or benchmark.

• Every file path, diff, commit, branch name, timestamp and terminal line: src/session.ts and its 12 dependents, the 10:01 / 10:02 / 10:03 collision, the 13-passed-1-failed test run, the fix/session-expiry branch, the handoff fields and the audit trail rows.
• The state meters in the ceremony chapter, which show relative weight rather than any measured overhead, and carry no numbers.
• The comparison bars for confusion against ceremony, and for reconstructing a run with one agent against several. Both are arguments drawn as diagrams, not data.

Not verified

• OMG is pre 1.0 and the project has published no tagged stable release, so every property stated about it here comes from what the project documents about itself rather than from an independent evaluation or from running it.
• No measurement of OMG in use exists to cite. Nothing on screen claims one.
• The central argument, that multi agent coding fails on missing shared state rather than on model capability, is the video’s own case. It is not sourced to a study, and no shot presents it as one.