Diagnose a missing handoff before changing frameworks
A fictional botanical garden updates visitor guides through source extraction, drafting and review. In a trial, a researcher reports completion while the writer receives an old file, and a reviewer merges feedback from another run. The immediate problems are task identity, communication and acceptance, not framework fashion. This guide preserves the originals' splitting, messaging and review stages with a reproducible handoff example. Central coordination, subagent calls and state routing can all be useful patterns. Multiple windows or an agent used as a tool should not be dismissed solely by form. Test a small task before adding complexity. No real team was launched, cross-session message sent or guide published for this article.
Stage one: define task cards before dispatch
Specify plants, approved source revision, output languages, page scope and unresolved information. Split extraction, drafting, review and merge with dependencies. Independent source groups may run concurrently; drafting waits for accepted facts and merge accepts only reviewed artifacts from the same run. Do not fragment a goal into chats that cannot be assessed independently. A task card includes run and task IDs, input revision, scope, tools, artifact location, acceptance and failure return. Researchers cannot modify the website incidentally, and writers cannot expand access to obtain extra material. The coordinator owns state and each artifact has one clear writer; other roles return comments or patches. Supply necessary context with provenance and restrictions. Do not assume a subagent remembers the parent, or assume it must be stateless. Inheritance and persistence depend on the environment and configuration. Full history can increase cost and conflict, while excessive compression loses requirements. Use explicit cards and readable evidence instead of relying on a persona name. Dry-run dispatch by checking the worker's interpretation of revision, goal and exclusions. This can be a structured contract check rather than another user approval for every ordinary action. Continue within existing authorization, and return a missing-input or permission state instead of quietly changing source or scope. Deliver a dependency map, cards, input snapshot and state table. Begin with a few bounded steps; task needs determine role count, not a fixed fleet of persistent sessions. Framework scheduling does not design the contract for the team.
Stage two: separate message acknowledgement from completion
Use an envelope with sender, recipient, run, task, input revision, message type, artifact reference and necessary summary. Types may include assignment, acknowledgement, progress, result, blocked, revision and cancellation. An acknowledgement means receipt, not completed work. A result still needs an existing, current artifact that meets the contract. Handle duplicates, delay and reordering. A late result for an earlier input must be marked stale rather than overwriting a current file. Repeated delivery must not merge an artifact twice. Establish message identifiers and valid state transitions with clear update ownership. Polling needs bounded intervals and timeout; event-driven designs still need duplicate and missed-event handling. Artifact references with concise summaries often avoid broadcasting full conversation history, but recipients must be able to read them and uncertainty must remain visible. Shared storage requires ownership and version control. A session label identifies work; it is not security isolation or account permission. On timeout, distinguish unreceived, running, failed and unknown outcomes. Inspect task identity and artifacts before redispatch, particularly for writes, publication or sending. Cancellation identifies the task and disposition of subsequent output so a late result cannot enter the final delivery. Recovery needs linked state and message evidence rather than inference from the latest chat. The originals name sessions_spawn and sessions_send. These are environment-specific tools, not universal interfaces across agents and frameworks. Verify real dispatch, status, result retrieval and cancellation support. A skill's description does not create a communication channel, and similarly named tasks, sessions and checkpoints should not be mapped as identical concepts without inspection. Deliver the message contract, state transitions, stale and duplicate rules, timeout and cancellation policy and synthetic trial records. Receiving one ordinary result is only a beginning. Test delayed old output, duplicate assignment, unreadable artifacts and worker failure to ensure the workflow does not advance incorrectly.
Stage three: review evidence instead of trusting role labels
Review input revision and scope before structure and claims. Every flowering-date statement must trace to supplied evidence. Missing sources remain gaps rather than being filled from model memory. Deterministic checks cover fields, identifiers and reference completeness; semantic review asks whether a source supports the statement. Avoid multiple roles producing the same vague score. A pass, revise or reject decision needs location, standard, evidence and retest. Rejection need not mean rewriting everything. Revision returns a new artifact and change explanation. Set round and budget limits and return unresolved questions for an owner rather than allowing endless debate to produce a pass. Cross-review can surface different issues, but six agents are not automatically six independent sources. Shared summaries, prompts and evidence can produce shared mistakes. Votes, confidence and self-evaluation are not factual correctness. Significant decisions, conflicting sources and irreproducible claims return to responsible people or reliable material. Writing Guidelines fits editing the technical handoff specification. Its public listing, pinned Vercel source and MIT license have been checked. Supply a redacted draft and audience to improve terminology, prerequisites and failure instructions. It does not create sessions, review botanical facts or provide orchestration. The association is restricted to documentation. Before merge, confirm a shared input snapshot, readable references, accepted revisions and retained unknowns. Deliver the draft, source table, review and change record, not merely a claim that the team finished. Publication is separate and is not automated in this example.
An end-to-end trial
Prepare three public or synthetic plant records, including a missing date and an older description. Dispatch cards, deliberately delay an old result and redeliver one result. Check that the writer uses current material, review identifies missing evidence and merge avoids duplication. Cancel an unfinished task and ensure late output does not join the final draft. A bounded request is: “Design, or verify in the authorized test environment, a garden handoff with identity, versions, acknowledgements, duplicates and cancellation. Review with located evidence. Do not create external accounts, widen file access or publish. Separate execution from unexecuted plans.” Compare sequential and split approaches on the same fixture before adding roles.
FAQ and troubleshooting
Which framework first? Start with state, tools, communication and operating constraints, then test a common fixture. Do not equate one framework with a single diagram pattern. Are manual windows useless? Human dispatch can work; automation addresses different overhead and is not inevitably superior. Receipt without a result requires task and artifact inspection rather than treating acknowledgement as acceptance. An artifact exists but is wrong: inspect input snapshot, output revision, writer ownership and review. Conflicting reviewers need their evidence and standards compared rather than a vote on facts. Cross-client portability requires verifying format, tools and features; licensing, model use and service costs also vary. No verified SkillPI orchestration or communication candidate was obtained, so those associations are not invented. Reassess future packages by actual dispatch, messaging or review purpose, public listing, source revision, license and prerequisites. The writing association is not the whole collaboration system.
Sources and maintenance
The paired originals cover decomposition, communication, review and FAQ. These remain with original garden and out-of-order-message fixtures. References are LangChain multi-agent and subagent documentation, CrewAI task-output validation and the writing package's pinned source. Popularity, universal tool names, absolute compatibility and automatic-review guarantees are omitted. The two matching archive variants share a stable valid public article slug, separate from the Chinese production-recovery variant.