AnswersPublished12 sources
AI Agents Need a Workflow, Not a Bigger Cast of Characters
The blueprint treats Analyst, Architect, Coder and Supervisor as stages in one workflow—not as four separate agent systems. A task can move downstream only when required machine checks pass and an independent reviewer approves the same candidate.
Published byImages generated with GPT Image 2
Research answer
Key takeaways
- The blueprint treats Analyst, Architect, Coder and Supervisor as stages in one workflow—not as four separate agent systems.
- A task can move downstream only when required machine checks pass and an independent reviewer approves the same candidate.
- The document is a draft specification, not evidence that its tools, safeguards or host integrations have been implemented.
![[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/700B227E2393C401EE05/54741874CB99CCCF8AA0.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":"以下规范将任务单元作为端到端业务流的唯一流转契约，采用“机器检查通过＋独立审查签章”控制下游释放。交付范围为工程规范草案；未修改工作区、实施代码、执行测试或迁移现有资产。\n\n BMAD Solo 业务流驱动的 Agent 文档与任务治理落地规范\n\n版本：0.1。日期：2026 10 27。状态：DRAFT. 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
AI agent systems do not become reliable simply by adding more specialized roles. A recent paper listing for AutoSwarm highlights a familiar risk: dividing work among agents does not automatically produce complementary contributions, and dependencies can allow local errors to become task failures. 1
One proposed answer is to focus less on the number of agents and more on how work moves between them. A draft governance framework for BMAD-Solo organizes tasks as a single business workflow, with two broad modes—Plan and Act—and four responsibilities: Analyst, Architect, Coder and Supervisor.
One workflow, four responsibilities
In the proposal, the Analyst clarifies the problem, requirements and unknowns. The Architect turns those into a task contract: what must be delivered, which files are in scope, how success will be checked and what should happen if a check fails. The Coder implements the approved work. The Supervisor reviews the resulting candidate and its evidence.
These are workflow stages, not four independent agent systems or separate stores of task status. The aim is to keep the work tied to one authoritative task contract rather than relying on scattered notes or chat history.
That contract would record the task’s goal, acceptance criteria, inputs, dependencies, allowed changes, verification plan and current status. A separate pointer could identify the active task, but it would not become another place to maintain the task’s progress. The draft also recommends preserving historical task records and evidence rather than treating an old output folder as disposable.
Two gates before work moves downstream
The central safeguard is a pair of checks. First, a trusted verification process must run the required tests and produce a result tied to the exact candidate being reviewed. Second, an authorized reviewer who did not implement that candidate must approve it.
Both gates must apply to the same version of the work. A passing test result for an earlier version would not be enough, nor would a reviewer’s approval if the files changed afterward. Missing, stale or unclear evidence should block release rather than be treated as a pass.
The distinction matters because an AI system can prepare code and even draft a review, but that does not make its own approval independent. Under the proposal, an implementation agent may assist with the review, while a separate authorized person or genuinely independent review service provides the final sign-off.
Keep handoffs focused—and preserve uncertainty
Each stage should pass on only what the next one needs: the Analyst’s findings and open questions; the Architect’s approved requirements and checks; the Coder’s frozen candidate, changes and test evidence; and the Supervisor’s decision, limitations and guidance for downstream use.
The framework also warns against presenting unknowns as successes. If a required check was not run, a result cannot be parsed or the reviewer’s independence is unverified, the task should be marked as blocked or lacking sufficient evidence—not accepted.
This emphasis on putting quality controls inside the workflow has a parallel in Huawei’s stated approach of building quality requirements into end-to-end business processes. 7 The proposal applies a similar principle to AI-assisted software work: quality should be checked at handoffs, not left to a final claim that the work is done.
A blueprint, not a deployed system
The document is explicit about its limits. It is a draft engineering specification, not proof that a trusted test runner, protected approval service or hard enforcement in Roo Code or Cline already exists. It says those capabilities would need to be implemented and tested—including checks that reject self-approval, stale results and changed candidates.
That caveat is the point: prompts can describe a process, but they cannot by themselves prove that permissions are isolated or that a release gate cannot be bypassed. The proposed framework therefore treats implementation evidence as part of the design, not as a detail to assume away.
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 "AI Agents Need a Workflow, Not a Bigger Cast of Characters"?
The blueprint treats Analyst, Architect, Coder and Supervisor as stages in one workflow—not as four separate agent systems.
What are the key points to validate first?
The blueprint treats Analyst, Architect, Coder and Supervisor as stages in one workflow—not as four separate agent systems. A task can move downstream only when required machine checks pass and an independent reviewer approves the same candidate.
What should I do next in practice?
The document is a draft specification, not evidence that its tools, safeguards or host integrations have been implemented.