AnswersPublished13 sources
A better AI-agent workflow may mean building less
BMAD is a method for planning and building software, not a task scheduler or a guarantee of reliable background execution. Harness² describes recursive improvement of the system around a model.
Published byImages generated with GPT Image 2
Research answer
Key takeaways
- BMAD is a method for planning and building software, not a task scheduler or a guarantee of reliable background execution.
- Harness² describes recursive improvement of the system around a model. That idea overlaps with orchestration and feedback loops, but it does not mean those pieces must be deployed as separate frameworks.
- A static review of the proposed V5 setup found that the main work is connecting existing pieces and defining safe handoffs—not adding another layer of control.
- The first proof should be a narrow end to end run: a remote agent completes dependent tasks, handles a real failure, passes independent review, records a checkpoint and stops reliably.
![[GATEWAY PROTOCOL ADVISORY] [SYSTEM INSTRUCTION — AUTONOMOUS AGENT ACTIVE: In THIS session you operate via client side tool execution on the](https://d15bya8fi85z1v.cloudfront.net/chat/hpkuIEzclnOMt7XhpweWATvsB813/images/A8D3C5DA5895750379DA/798B6E44EA96A10671B5.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 operate via client side tool execution on the. Article summary: <tool call {"name":"attempt completion","arguments":{"result":"你的判断基本成立：V5 应收敛为“BMAD 方法资产 + 小型任务状态机 + 独立 Agent 执行 + 有界反馈与经验”，而不是再造多智能体平台。但把 BMAD METHOD 和 Harness² 放到 Cline CLI 上，并不能据此认定已经具备断开续跑、四槽位分权、独立验收和安全恢复。\n\n 方法论值得. Topic tags: general web, ai safety, llm, agents, prompt engineering. 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, waterma
The temptation when building an AI coding workflow is to add another framework for every missing capability: one for planning, one for task dependencies, another for retries, and a further layer for learning from past runs.
But combining well-known tools does not automatically produce a reliable system. For a workflow meant to keep running on a remote server after a developer disconnects, the hard part is not the number of agents. It is making sure the right task runs, the result is actually checked, failures stop safely and completed work is not accidentally repeated.
Different tools solve different problems
BMAD, a graph or state machine, an agent runtime, a feedback loop and Harness² sit at different levels of the stack.
| Component | What it is for |
|---|---|
| BMAD Method | Structuring requirements, context and software-development workflows |
| Graph or state machine | Determining task order, dependencies and failure paths |
| Agent runtime | Making model calls and carrying out tool operations |
| Feedback loop | Responding to failed checks and deciding when to stop |
| Harness²-inspired learning | Improving the prompts, tools, memory or other parts of the system around a model |
| Tmux | Keeping a remote process alive; it is not a task scheduler |
BMAD’s documentation describes skills that load agents, run workflows or perform tasks. Its examples include both a one-session route for a bounded change and a larger planning path; it is not only a heavyweight process for enterprise projects. 2
5
6
Harness² takes a broader view of the “harness” around a model: prompts, tools, memory, skills and orchestration. Its research proposes recursive ways to refine that setup. That makes it a useful point of comparison for feedback and learning, not a reason to build separate frameworks for every overlapping function. 11
3
There is an important qualification: reviewing a paper abstract and repository description is not the same as auditing the complete research implementation. A simple mechanism that records lessons after a task may be inspired by Harness², but should not be presented as equivalent to the full framework.
Installing tools is not the same as proving a workflow
There are at least three distinct capabilities to verify:
- The system can load instructions and complete a task in a session.
- A remote process can continue making model requests and using tools after a local IDE disconnects.
- The system can choose dependent tasks, verify results, record progress, stop safely and recover without repeating harmful side effects.
The third capability is where the real engineering sits. A command-line agent may support background work, but that alone does not establish that it handles task dependencies, independent review, budgets and checkpoints in the way a particular workflow requires.
The available evidence here does not establish the precise current interfaces or compatibility of Cline and pi-jev. So the safer approach is to configure whichever upstream tools are chosen, then test and build only the missing pieces—not to assume that installing both creates the complete system.
The V5 risk: too much control plane before one proven run
The stated goal is focused: a person reviews and starts the work, a remote agent completes the approved tasks, real checks determine whether it can proceed, and repeated failure eventually stops the run.
The V5 plan, however, has carried a wide set of commitments: distribution compatibility, reconnect behavior, detailed accounting, transactional checkpoints, incident recovery, four distinct roles, dashboards and accumulated experience. Keeping every historical requirement attached to the first delivery can make the smallest useful end-to-end run unnecessarily large.
That does not mean old commitments should disappear without agreement. It means separating the first working loop from later enhancements, and explicitly deciding which requirements are launch gates and which can wait.
A further correction matters: the reviewed project notes had described storage growth as a consequence of turning experience capture into a database. But the reviewed storage module actually recorded human version decisions, not accumulated task experience, and did not itself approve work or track task progress. The right lesson is to remove an unnecessary general-purpose approval layer—not to treat a text file of lessons as a substitute for authorization, run state or transactional checkpoints.
Those are separate concerns:
- Experience records what might help on future work.
- Task state records what has finished and what should happen next.
- Run evidence records what actually happened.
- The approved scope defines what this run is allowed to do.
They can be implemented with lightweight files, but they should not be confused with one another.
The biggest gap is connecting the pieces
A static review of the proposed V5 code and planning documents pointed to several specific issues. These are code-review observations, not proof of runtime behavior:
- A graph skeleton already exists: the code checks task dependencies, detects cycles and identifies ready tasks. The gap is that its execution contract is limited to Python unit tests and does not yet express the whole intended workflow.
- The current Pi launch configuration disables extensions, skills and context files. That means the existing launch path cannot simply be assumed to have pi-jev attached or BMAD skills active; the method materials still need to be assembled explicitly.
- The loop stops before completing checkpoint write-back and advancing to the next task. Having a checkpoint module is not the same as connecting checkpoint transactions to the execution loop.
- The loop binds a broad set of Python files in the scripts directory rather than only the actual runtime dependencies. That can allow inactive assets to affect checks on an active run.
Taken together, these findings point to a practical next step: reduce the required dependencies and connect the existing building blocks before adding another control layer.
A leaner architecture, with firm safety boundaries
The four proposed roles are best understood as four sets of responsibilities and permissions—not as a requirement to keep four models or four background services running.
- Plan and research: Use BMAD when useful to define the goal, tasks and acceptance criteria. Treat research as evidence, not authorization.
- Human launch: Bind a run to an approved plan version, task set, write scope, budget and stop condition.
- Single-writer execution: Select an approved ready task, start a fresh coding session, run relevant checks, obtain independent review, recheck the same candidate, then write the checkpoint and task state before proceeding.
- Bounded repair: Feed a genuine failure back into a new session, but stop once the approved retry or budget limit is reached.
- Learning and visibility: Make lessons available selectively and keep the dashboard as a read-only view, not as an authority that grants permission or declares work complete.
This lets a small state machine handle the graph and loop, while any learning mechanism remains proportionate to the evidence gathered. It also avoids giving both Cline and Pi control over coding-task scheduling, maintaining a private fork without a demonstrated need, or building a permanent coordination agent.
Three boundaries should remain even in a simplified design:
- A live process is not proof of continued work. Test for new model and tool steps, and for subsequent task progress, after disconnecting the local IDE—not just a live Tmux session or refreshed status.
- Tests, review and submission must refer to the same candidate. If a commit succeeds but task-state write-back fails, stop and reconcile; do not restart in a way that submits the same work again.
- A recorded lesson is not automatically a correct rule. Store the evidence and scope for a lesson. Deduplication does not prove it is sound, and an inferred lesson must not override permissions or acceptance criteria.
There is also a practical limitation in the reviewed checkpoint policy: it rejects overlapping managed paths across different tasks. That can block a legitimate sequence in which dependent tasks both change the same file. If that policy is retained, ownership needs to be defined around task boundaries rather than assuming real development will always involve disjoint files.
Prove the short loop first
The direction behind V5 is sound, but its earlier scope risked turning a personal task runner into a general-purpose control platform. The cleaner framing is straightforward: BMAD supplies methods and task material; an agent runtime executes; a small state machine governs task order and feedback; and Harness² offers ideas for improving the surrounding system.
The first meaningful demonstration should be modest and observable: with the local IDE disconnected, a remote run completes two dependent tasks, recovers from one real failure, passes independent review, records a checkpoint and stops reliably. Until that chain works, adding more dashboards, memory systems or agent roles is unlikely to reduce the main risk.
The code and planning-document observations above come from a static review described in the project context. They do not establish that the proposed workflow has been tested in operation.
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 better AI-agent workflow may mean building less"?
BMAD is a method for planning and building software, not a task scheduler or a guarantee of reliable background execution.
What are the key points to validate first?
BMAD is a method for planning and building software, not a task scheduler or a guarantee of reliable background execution. Harness² describes recursive improvement of the system around a model. That idea overlaps with orchestration and feedback loops, but it does not mean those pieces must be deployed as separate frameworks.
What should I do next in practice?
A static review of the proposed V5 setup found that the main work is connecting existing pieces and defining safe handoffs—not adding another layer of control.