Map dependencies before deciding how many agents to use
A fictional community museum is preparing bilingual exhibit labels. Source extraction, drafting and webpage checks appear separable, but faster writing only amplifies errors if the facts are not established. Multi-agent value comes from responsibilities, input contracts and acceptance rather than giving chat windows job titles. This guide preserves team design, orchestration and framework selection through one original label-delivery example. Establish a single-agent baseline using the same materials and record quality, time, calls and human rework. Add roles only when a task can be separated and expected gains justify handoff overhead. No agents were launched, frameworks installed or labels published for this article.
Scenario one: define roles through deliverables
The museum owner sets scope, approved sources and final publication. A coordinator owns task state and versions; a source role extracts only assigned evidence; a writer drafts from the accepted fact table; a reviewer checks claims and requirements. These need not be separate agents at first. One agent can perform sequential steps before responsibilities are split. Titles do not grant tools, authority or competence. For every task, specify inputs, outputs, dependencies, allowed tools, file scope, acceptance and failure return. The source role delivers statements with dates, citations and unresolved questions. The writer identifies the fact-table revision. The reviewer returns located issues instead of an overall compliment. Agreement between agents cannot turn an unsupported statement into a fact. Separate independent work from dependent work. Source extraction over different assigned records may run concurrently; drafting waits for accepted facts and browser acceptance waits for an actual page. Provide sufficient provenance without sharing every file and chat by default. Summarized handoffs must retain uncertainty rather than converting tentative observations into established truth. Separate contexts are not security isolation. Roles may still share a directory, credentials or network. Assign write ownership and use appropriate work copies and a merge owner when needed. A directory boundary alone does not constrain a helper script's access to host data. Implement actual execution boundaries according to the environment. Deliver a role table, dependency map, permission inventory and sample accepted handoff. Human responsibility concerns meaningful decisions, not repeatedly approving every reversible step already in scope. Specify which outcomes need owner review and which ordinary processing may continue.
Scenario two: orchestrate state, not conversational enthusiasm
Use explicit task states such as waiting for input, ready, running, awaiting review, revision required, complete and failed. A handoff includes task ID, input revision, artifact location, conclusion, unknowns and next action. “Finished” in a chat is not a machine-readable acceptance condition. The coordinator updates state from evidence rather than self-assessment. The label flow can accept a fact table, prepare a draft and terminology reference in parallel, merge them for review and then deliver to the owner. Source or terminology conflicts return to a responsible step instead of initiating an unlimited agent debate. Set revision, time and cost limits and return unresolved issues when exhausted. Persist intermediate state so interruption does not erase knowledge of completed steps. Persistence alone does not make external actions exactly once. Before retries, distinguish pure generation, file writes and publication. Sending, posting and reservations require separate side-effect control and deduplication. When a timeout leaves the outcome unknown, inspect the actual result before repeating it. Layer quality gates. Required fields and format can be checked mechanically; claims need source comparison; cultural expression and publication need responsible judgment. A reviewer should identify source location, issue and proposed correction, including checks not performed. A second model is not automatically an independent audit if it shares the same assumptions. Writing Guidelines fits clarity and structure review of the technical handoff document. Its public listing, pinned Vercel source and MIT license have been checked. Supply a redacted procedure and audience. It does not start agents, provide persistence or scheduling, or verify museum facts. Its association is deliberately limited to documenting the workflow. Test missing evidence, review rejection, execution failure and duplicate submission. A successful path is insufficient; recovery and refusal must stop publication appropriately. Deliver state records, handoff artifacts, review results, failure handling and acceptance rather than a transcript that merely resembles collaboration.
Scenario three: compare frameworks on one fixture
List necessary conditions before names: implementation language, tools, state and recovery, human intervention, logs, deployment constraints and maintenance skills. The Chinese original emphasizes graph and role orchestration, while the English original also lists Microsoft Agent Framework and other candidates. Treat them as options to assess rather than a bundle to install. Include an existing sequential script as a baseline. LangGraph offers mechanisms for stateful orchestration, persistence and human intervention. CrewAI offers a role- and workflow-oriented route. Microsoft Agent Framework's public implementation covers agents and workflows in its supported languages. Verify the selected versions, interfaces, licenses and operational prerequisites upstream. Slogans about graphs being complex and roles being simple do not settle project fit. Use the same label fixture in a controlled comparison. Check retained facts, located errors, rejection, interrupted recovery, repeated side effects, per-run cost and human rework. Record model, tools, versions, concurrency and retry policy. Different inputs and budgets do not establish a fair comparison. Without measurements, present an experiment plan rather than invented speed gains. Explain selection and rejection through migration, state storage, dependencies, observability and team learning. A demonstration feature does not establish that the team can operate it reliably. Start with a bounded pilot and keep a sequential fallback. Low task volume or strong dependencies may favor fewer agents or no additional framework. The current SkillPI framework search did not identify a verified orchestration candidate, so no framework skill link is invented. The open source projects are external references and the writing association supports documentation only. Reassess future packages by pinned source, license, prerequisites and exact step rather than name similarity.
A practical design request
“Using synthetic label records, establish a single-agent baseline plan, then specify roles, deliverables, dependencies and permissions. Parallelize only independent work. Define revisioned handoffs, review gates, failures and retry limits. Compare an existing script with candidate frameworks on the same fixture. Do not connect external accounts or publish. Distinguish a design from actual execution records.” This avoids interpreting a request for task decomposition as unlimited delegation and action.
FAQ and troubleshooting
Are multiple chat windows collaboration? They can be manually coordinated; effectiveness depends on responsibility and handoff, not window count. Automated orchestration adds state and scheduling along with maintenance cost. Are more agents faster? Not necessarily: waiting, duplicated context and review can increase time. File conflicts require ownership and merge control rather than a claim of context isolation. Why can every agent agree and still be wrong? They may share a bad source or inference. Return to evidence, not votes. Can an interrupted run be restarted immediately? Inspect state and external outcomes first. Does installing a skill supply orchestration? It may only supply a procedure; framework, runtime, tools and access have separate prerequisites. Costs and licenses also need separate review.
Sources and maintenance
Both originals cover team construction, workflow orchestration, framework comparison and FAQ. These remain with original label fixtures and failure paths. References include LangGraph, CrewAI, Microsoft Agent Framework and the linked writing package's pinned source. Popularity, automatic-takeover, absolute-reliability and free-use guarantees are omitted. Refresh technical versions and associations while preserving the public URL and language pairing. This guide does not claim that a real multi-agent system was run.