What works today, and what does not

Arcadia governs its own development — these tags describe the last reviewed snapshot. Working means demonstrated in the dogfooded system; it does not imply general availability.

Snapshot reviewed 2026-11-09. What changed →

• Plans, actions and a work pointer

  Governed documents with one active plan and one current action.

  How Project and Plan documents record the active work.
  Working

• Arcadia Field Notes opens Canyon Pass

  Field Notes now has a moving Martian canyon that follows the site’s day and night cycle. Readers who prefer reduced motion get a still backdrop. The November 9 post follows a real test run from a confusing instruction to a check before launch.

  Where to read Field Notes and how Canyon Pass works.What Arcadia checks before an agent starts.
  Working

• Earlier warnings for confusing run instructions

  Before preparing a run’s start card, Arcadia warns about ambiguous test filenames and rejects certain invisible characters in the instructions. The check passed offline tests and CI. Its effect on live retries is still unmeasured.

  The original failure, the tested fix, and the limits of the evidence.The checks Arcadia runs before it prepares a start card.
  In progress

• Agent session management

  Isolated worktrees per action, across Codex, Claude Code and OpenCode.

  How Arcadia hands an Action to Codex, Claude Code, or OpenCode.
  Working

• Decisions and approval gates

  Consequential calls are recorded and wait for a human answer.

  Arcadia’s rules for human approval.
  Working

• Completion evidence

  Acceptance criteria are checked before an action can be marked done.

  The evidence required to settle a completed Action.
  Working

• CI and review-finding repair

  A failing check or a reviewer's finding gets fixed and re-validated automatically, up to a bounded number of attempts, before it reaches you.

  Tests for review findings, repairs, and retry limits.
  Working

• Multi-project portfolio view

  One surface across every project you are running.

  Flight Deck and the local portfolio views.
  In progress

• Install on someone else's machine

  Cold-start setup for a person who is not the author. Early access waits on this.

  Current local setup steps and prerequisites.
  In progress

• Hosted Mission Control

  Today it runs locally, on your machine and your repositories.

  Current local installation scope.
  Exploring

• Cost-aware model selection

  Spend the cheapest model that can actually do the work.

  How Arcadia chooses model tiers and routes.
  Planned

• Idempotent Ask intake

  Immutable Ask intake and two-phase preview and settlement.

  Tests for Agent Ask previews and settlement.
  Working

• Next-action clarification

  Clarification produces one concrete next action or one question.

  The clarification loop and the single question it can raise.
  Working

• Independent pull-request review

  Independent review is required before merge.

  Independent review and merge requirements.
  Working

• Peer-watch building blocks

  Trailer parsing, classification and operator nudges exist; a running peer watcher does not.

  The peer-watch contract and its operating limits.
  In progress

• One to-do list for what needs your attention

  Run arcadia todo to see pending Decisions, Agent Asks and review items together, with blockers first. Arcadia builds the list from existing records each time you open it.

  The read-only to-do command and its inputs.Tests for blockers, ordering and duplicate items.
  Working

• A check that the next Action is ready to start

  When Arcadia clarifies a request, it checks that the next step is concrete and has an observable finish. A separate grading pass can send back one question for the missing detail.

  The completed grading work and its acceptance criteria.How clarification checks the done-condition and handles an open question.
  Working

• Free-text input coverage

  Capture of Decision and review replies with an input-coverage report.

  The scope of free-text reply capture and coverage reporting.
  Planned

• Standard harness protocol

  A written standard-harness procedure and session protocol.

  The written standard-harness procedure.
  Planned

• Workflow registration and projections

  GitHub and Flight Deck projections with automatic workflow registration are deferred.

  The deferred workflow work and what would bring it back.
  Deferred

Arcadia is real software in active development. It is not finished, it has no customers or measured productivity outcomes yet, and early users help decide what it becomes.

What early access would mean, and when it opens →