AnswersPublished12 sources
A blueprint for keeping AI-assisted work accountable
The proposal keeps each task’s contract, status and acceptance evidence in one authoritative record rather than scattering them across role specific documents. One agent can move through Analyst, Architect, Coder and Supervisor duties, but changing roles does not grant new permissions or make its review independent.
Published byImages generated with GPT Image 2
Research answer
Key takeaways
- The proposal keeps each task’s contract, status and acceptance evidence in one authoritative record rather than scattering them across role specific documents.
- One agent can move through Analyst, Architect, Coder and Supervisor duties, but changing roles does not grant new permissions or make its review independent.
- Release requires both passing machine checks tied to the exact candidate and a valid sign off from someone who did not implement it.
- The document is a proposed engineering specification—not evidence that the workflow, code or host tool integrations have been built or tested.
![[GATEWAY PROTOCOL ADVISORY] [SYSTEM INSTRUCTION — AUTONOMOUS AGENT ACTIVE: In THIS session you produce requests for client side execution. W](https://d15bya8fi85z1v.cloudfront.net/chat/GwlKmI7QinNfXevZji84RFDPF1F3/images/3691292DB166B5BB21C7/D32C36C0DF02CF52EB04.png)
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: [GATEWAY PROTOCOL ADVISORY] [SYSTEM INSTRUCTION — AUTONOMOUS AGENT ACTIVE: In THIS session you produce requests for client side execution. W. Article summary: <tool call {"name":"attempt completion","arguments":{"result":"本草案采用“一个任务工件贯穿业务流、两个基础动作承载四种职责、机器检查与独立签章共同放行”的治理模型。交付仅为工程规范设计，不实施代码、不迁移现有资产，也不把提示词切换视为独立审查或运行时权限隔离。\n\n 任务契约与生命周期统一归入 specs/tasks/，不按角色建立文档孤岛。\n 同一 Agent 可以切. Topic tags: general web, llm, agents, ai, automation. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts wit
The central idea in a new BMAD-Solo governance blueprint is simple: an AI agent may take on several jobs, but its change of role should never be mistaken for a change of authority.
The draft lays out a single task record that follows work from planning through implementation and review. It proposes two kinds of release checks—machine verification and independent human approval—and stresses that neither a prompt change nor a “done” message can replace them.
This is a design proposal, not a report on a deployed system. Its author explicitly treats the paths, templates, gates and Roo Code/Cline mappings as proposed assets, not as features already created or validated.
One task, four responsibilities
The workflow groups four responsibilities into two broad stages. During Plan, an Analyst clarifies the business goal, constraints and acceptance criteria; an Architect turns those into interfaces, deliverables, dependencies and verification checks. During Act, a Coder implements the approved work, then a Supervisor organizes verification and review.
That division is about the work to be done—not four separate identities. A single agent may switch between the responsibilities, but the switch does not grant new permissions. In particular, an agent that wrote a change may help prepare a review or look for defects, but it cannot provide the independent sign-off on its own work.
That distinction matters in systems built around specialized agents. The AutoSwarm paper notes that dividing work among agents does not automatically produce complementary contributions, and that dependencies can allow local errors to affect the wider task.1 The proposed framework responds with explicit task ownership, evidence requirements and approval boundaries. It does not claim that those measures have been experimentally proven to solve the problem.
Passing tests is only half the gate
The blueprint’s release rule combines two conditions: the required machine checks must pass for the current version of the candidate, and an independent reviewer must approve that same candidate. Missing or stale evidence is not a pass. Nor is a successful test run enough if the review identifies a business or safety issue that the tests cannot assess.
For the machine side, the proposal calls for a verifiable record linking the task contract, candidate, test results and execution environment. A missing test, timeout, failed check or unknown result blocks release. The review side is similarly tied to the candidate: if code or relevant configuration changes after approval, the earlier evidence and sign-off may no longer apply.
The default model is an independent human reviewer. A separate automated review service could be considered, the draft says, but only after its identity, permissions, inputs and failure conditions are explicitly governed. A new session, a different model or a different role label does not by itself establish independence.
Keep the record in one place
Rather than maintaining separate task ledgers for each role, the proposal gives each task one authoritative contract and current status. The record would hold its goal, scope, acceptance criteria, dependencies and verification plan, alongside references to evidence and approvals. A separate “active task” pointer would identify the current focus without duplicating its stage or acceptance state.
That structure is intended to prevent conflicting versions of “what is done” from spreading across chat summaries, host-tool to-do lists and project documents. Domain documentation would describe established rules and interfaces; it would not become a second task-status ledger. The proposal also distinguishes acceptance from authorization to merge, actual merging, and archiving—separate events that need separate evidence and permissions.
The underlying process-first rationale has a parallel in Huawei’s public discussion of quality management, which calls for quality requirements to be built into business processes.7 That is a point of context, not proof that this particular AI-agent design will work in practice.
Tool settings are not security controls
The draft also cautions against treating configuration as enforcement. A Roo or Cline prompt can guide an agent’s behavior, but a static rule file does not prove that a host loaded the rule, restricted the agent’s tools or blocked an invalid release. Those behaviors would have to be tested in the actual runtime.
Likewise, calling a folder a “sandbox” does not make it one. The proposal says isolation needs to come from real limits on file access, network access, credentials and processes. If the environment cannot provide those controls, the limitation should be disclosed rather than described as a guarantee.
The same principle governs failure and archiving: preserve the failed candidate and its evidence, do not overwrite user changes, and do not label unfinished work complete simply because it has been filed away.
What the proposal does—and does not—establish
The blueprint offers a detailed target design: one task-centered record, role changes without privilege changes, machine checks bound to a specific candidate, and independent approval before release. It also acknowledges that the necessary validators, runtime controls and integrations are not established by the document itself.
That makes the key question practical rather than rhetorical: can the chosen tools enforce these boundaries before work advances, and can a reviewer verify the evidence independently? Until those controls are implemented and tested, the framework remains a specification—not a guarantee of safe or reliable AI-assisted development.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "A blueprint for keeping AI-assisted work accountable"?
The proposal keeps each task’s contract, status and acceptance evidence in one authoritative record rather than scattering them across role specific documents.
What are the key points to validate first?
The proposal keeps each task’s contract, status and acceptance evidence in one authoritative record rather than scattering them across role specific documents. One agent can move through Analyst, Architect, Coder and Supervisor duties, but changing roles does not grant new permissions or make its review independent.
What should I do next in practice?
Release requires both passing machine checks tied to the exact candidate and a valid sign off from someone who did not implement it.