What Went Wrong in Designing an AI Organization Harness

In Activity 0004, sessions that only named a role in the prompt had not applied the Custom Agent configuration; launches that passed the role ID through agent_type confirmed the configured profile.

In Activity 0005, some agents still identified themselves with a different role and nickname in their responses even though the invocation role ID, UI nickname, model, and reasoning effort matched the configuration.

In Activity 0072, resuming a session did not restore previously completed subagents to the live tree.

We first verify the schema, accepted characters, tool API, and runtime metadata, and only then design organizational rules.

We separate role IDs, UI nicknames, and self-identification, and verify invocation through runtime information.

In Activity 0034, the AI responded to a request to reduce approval fatigue for routine edits by proposing a broad permission profile that the CEO withdrew as inconsistent with the request.

In Activity 0048, a strict completion-receipt proposal would also have blocked paused, cancelled, and parallel work, and the CEO did not adopt it. In Activity 0051, a session-inspection helper for a simple conversation check became too complex, and the work was stopped.

In Activity 0054, even a four-stage lifecycle refined through Ritsu's reviews successively blocked readying, merging, and cancellation; closing the pull request and Issue finally required a temporary repair to the helper itself.

Activity 0055 removed the custom hooks and wrappers and returned execution control to the standard sandbox and approval model. Activity 0220 and 0223 later established independent review, CEO merge approval, Issue-to-PR relations, and branch-to-PR relations as operating rules.

Activity 0214 combined corporate information and the product experience into one Web build. Activity 0215 froze the former public site, and Activity 0217 removed work records and snapshot machinery that were no longer needed. Activity 0219 moved agent definitions and routing from outside the project into the same repository as the source they govern.

We use the standard sandbox and approval model as the foundation, while the project retains mutation ownership, Issue-to-PR relations, independent review, and human merge approval.

Current rules, agent definitions, specifications, design, and their execution target are kept close within the project that uses them.

We protect boundaries around secrets, destructive actions, publication, and merging, and inspect actual scope and side effects when an unexpected state occurs.

Activity 0007 and 0040 centralized GitHub connection handling, CLI fallback, authentication checks, and role routing in repository-local wrappers and a project Skill.

Activity 0055 removed the custom hooks and wrappers and returned external operations to scoped direct execution followed by verification of actual state.

We confirmed that a multi-stage delegation chain could launch grandchild agents, as in a sequence from Mako to Ryoma, Manabu, and Ikumi.

However, Mako could not see that it should wait for a grandchild agent's handoff and sometimes assigned other subagents work that could conflict with the grandchild's task. Mako also ended its own session before the grandchild returned, leaving completed handoffs uncollected.

Resuming did not automatically restore completed agents to the live tree.

Activity 0024 split prose expression into a Scenario Writer role from a role that had owned narrative structure and writing.

Activity 0033 identified an unclear boundary between meaning design and Writing, together with duplicated transfer and verification of constraints, as reasons for reorganizing the roles.

We currently have no writing-only agent; one role owns the work from Narrative Design through Writing.

We define roles around coherent responsibility for an outcome and keep one meaning with one role for as long as necessary.

Creation and independent review remain separate.

Unless explicitly instructed to remove superseded text, AI agents tended to preserve the existing wording and append strikethroughs or negations for removed specifications. As a result, current specifications and past work records accumulated in the same document.

Successive revisions eventually negated earlier negations. The documents grew, the currently valid instruction became difficult for humans to identify, and the obsolete wording became noise that led AI agents to consult old statements or fail to follow current rules.

Copying the same source into handoffs and snapshots created additional update paths and made the current judgment even harder to identify.

We assume that AI agents may treat deletion as risky and choose safety-leaning additions instead. A specification therefore explicitly states that it is the master of current system behavior, not a work record. Update instructions also require replacing superseded specifications with currently valid content instead of retaining them as strikethroughs or negations.

We keep change history and historical facts separately in Activity, operating ledgers, and archives, referencing them rather than copying them into current specifications.

This map presents only the current interpretation, while the underlying historical facts remain available in their Activity records.

Narrative templates and historical review snapshots existed, but there was no single current version for CEO confirmation, downstream design, and publication-boundary decisions. Earlier reviewed versions could reintroduce decisions that had changed or wording that had never been approved.

Even after page copy was updated, the material handed to implementers still contained old wording and old rules about what could be published. Copying the source content into that material then left the same information in two places, making it unclear which one was current.

The current Narrative master separates fact, interpretation, hypothesis, and fiction. The Brand Bible, Narrative Guidelines, Fact Boundary, and Key Terms assign responsibility for Why, What, voice, fact boundaries, and key terms.

Downstream outputs such as About and Home align with the same sources and reference rather than copy their content.

New brand judgments absent from the sources return for human confirmation.