Agents

A workflow,
defined as data.

An agent is a JSON-defined production workflow — the single, reviewable source of truth between authoring and execution. Its runtime selects the engine; dynamic_subagents only controls fan-out. New production work becomes a new agent on the same runtime, never a new tool.

1

Structure

Anatomy of an agent

Agent definition · ch_<agent>.json IDENTITY & ROUTING id · name · runtime · category · enabled routing — how a ticket reaches this agent generic · unity · offer · data · engineering GROUNDING sources · retrieval · spec · coverage path_map · path_rules every value bound to a real source — never invented STAGES (the pipeline spine) ObserveTransform ValidateDecide Approve type · modeexecution_mode deterministicconfidence human gate configmanual·AI·auto gatepolicy deployer OUTPUTS & PROVENANCE validation · confidence thresholds · outputs · evidence studio — canvas layout (round-trips losslessly) · created_at / updated_at

Real top-level fields: id · name · runtime · dynamic_subagents · routing · category · enabled · path_rules · unity · validation · instructions · examples · stages · path_map · sources · spec · coverage · retrieval · studio. Each stage carries id · label · type · mode · execution_mode · config.

2

Live today

The agents in production

The defined agent roster by workspace. Live in production today: 3 Tech Art agents. Design's character setup and the Dev engineering / bug-fix engine are in development, not yet live; the other listed agents are defined and rolling out onto the same runtime as each contract is ready.

Ability (generic · 13 stages)

Ability-icon art: Photoshop export, sprite wiring, metadata, Git → the gotm client.

Flat Art Import (unity · 11 stages)

Zero-touch flat-art import: PSD → PNG, the _full sprite pivot, and ProtoDB sprite hookup via the live Unity bridge.

New Flat Art Setup (unity · 10 stages)

Stands up a new flat-art asset end to end in Unity — import, _full pivot, ProtoDB hookup — on a character branch off development.

PG Setup (generic · 7 stages)

Proving-Grounds setup workflow: assembles the required config and assets under the same governed contract.

Offers-DD Agent (offer · 8 stages)

Parallel Photoshop, dynamic dialog, offer packaging and LiveData sync → S3.

Offers-DD (offer · 13 stages)

Expanded offer production with artifact manifests, visual review, human sign-off and sealed publication.

DD Iconic Item (offer · 8 stages)

PSD/PSB → 764×1024 opaque PNG, confidence-gated, Content Bundles row → S3.

Design Character Setup (data · 12 stages)

Champion data/config setup into the gotm client — proto, shard, strings, sprite refs, content-strip: idempotent append on a new-character branch off development, dry-run by default.

Production Operations (7 stages)

Reads production planning data, builds release/work/resource intelligence, proposes Jira actions and performs governed writeback only after approval.

Offer and DD (offer · 7 stages)

QA-oriented offer and dynamic-dialog checks with the same governed runtime contract.

Engineering Bug Fix (engineering runtime · in development)

In development, not yet live. The governed bug-fix engine: a Bugsnag/JIRA crash advances a 17-state machine (observed → grounded → planned → patching → patched → reviewing → reviewed → validating → validated → awaiting_approval → approved → committed → draft_pr_created) over typed, trust-tagged artifacts. Patch runs under a CODE_FLOOR of 0.85; validation is blocked if any required check fails. It stops at a draft PR — never a merge or deploy.

Bugsnag Fix (dev · entry point · in development)

Focused Bugsnag ingestion that routes a crash into the governed engineering run above. In development alongside the engine.

Product workspace (shell ready)

Full workspace shell — pipeline, queue, approvals, knowledge — ready to commission domain agents through Studio when its contract is defined. No live agents yet.

Art workspace (shell ready)

Full workspace shell for the Art discipline, ready to commission domain agents through Studio. No live agents yet.

3

Operations

Lifecycle, testing & failure handling

Author

Compose & ground

Build in Studio; bind sources; set stages, policies, validators and confidence thresholds.

Certify

Lint · simulate · certify

Harness produces an explainable readiness certificate; golden baselines guard regressions.

Run

Preview → approve → execute

Dry-run shows reads/writes; a deployer approves; the runtime executes and streams progress.

Monitor

Observe & learn

Mission Control, telemetry, run journal and post-mortems record every run; outcomes feed future confidence.

Failure & retryOnly transient, idempotent connector calls retry automatically. Each batch job owns its artifact manifest and terminal outcome; one failed ticket must not prevent an independent ticket from running or cause its files to be committed.