Recommended Free Tools
A specialist can finish its work and still leave the workflow stuck: if its result is missing from the final message the supervisor receives, the parent agent has nothing usable to pass on. LangChain documents this as a common subagent failure mode. That is a design risk—not evidence that agents universally fail at “step 7.”
Who should own the final answer?
Decide this before splitting work across agents. OpenAI’s orchestration guidance frames the first design choice as deciding who owns the final user-facing answer at each branch.
![Claude AI for Beginners Bible: [5 in 1] The Ultimate Guide to Automate Your Work, Save Hours Every Week, and Use AI for Real-World Results](https://m.media-amazon.com/images/I/415+fSJacsL._SL160_.jpg)
As an Amazon Associate I earn from qualifying purchases.
Keep ownership with a manager
In a manager or agents-as-tools design, the manager retains responsibility for the user’s overall request. It calls specialists for bounded tasks, receives their results, and synthesizes the final response. This fits work where a central agent needs to apply shared guardrails or combine several contributions.
Hand control to a specialist when the branch itself matters
With a handoff, control moves to a specialist for the next branch of the conversation. Use this when the specialist should take over rather than merely return findings to a manager. Make the transfer and the context the specialist needs explicit.
#1 Best Overall
These patterns answer different ownership questions; neither is a reliability guarantee. A manager can lose a child’s result, while a handoff can leave a specialist without enough context to continue.
What should each agent be responsible for—and return?
Give each specialist a narrow assignment and make its handoff description concrete. Create a separate branch only when the work genuinely needs different instructions, tools, or policies. If the routing surface is hard to understand, first clarify the roles and their tool names, parameters, and descriptions rather than adding more agents.
For every child task, specify the deliverable the parent needs—not just the activity to perform. If the parent needs a decision, evidence, extracted fields, or a generated artifact, say so in the child’s instructions and require it in the final response. LangChain warns that a subagent can perform tool calls or reasoning yet omit the results from its final message; the supervisor sees that final output, not the child’s unreported work.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →When a free-form summary is too fragile, pass important fields through structured state in code. LangChain’s subagent guidance describes returning additional fields to the supervisor. Treat structured output as an explicit interface: identify required fields and check that they are present before the next step depends on them.
Should you use a supervisor, a router, or a fixed workflow?
A supervisor is a full agent that maintains context and can dynamically decide which subagents to call across multiple turns. A router is commonly a one-step classification and dispatch mechanism. If one agent with a few clear tools can complete the task, adding a supervisor may introduce unnecessary routing and state-transfer boundaries.
| Design | Who owns the answer? | Best fit | Main caution | Guidance |
|---|---|---|---|---|
| Manager / agents-as-tools | The manager retains ownership and synthesizes specialist results. | Bounded helper work, central synthesis, and shared guardrails. | The manager must receive and use the returned result. | OpenAI, “Orchestration and handoffs,” accessed 2026-10-07. |
| Handoffs / delegated ownership | The specialist owns the next branch response. | Cases where routing to a specialist is meaningful and that specialist should take over. | Keep branches and transferred context clear. | OpenAI, “Orchestration and handoffs,” accessed 2026-10-07. |
| Code-orchestrated workflow | The application determines the next step; an agent can perform bounded judgment tasks within it. | Fixed sequences, structured outputs, repeatable conditions, and explicit transitions. | The application must define the workflow and state handling. | OpenAI, “A practical guide to building agents” and “Running agents”; LangChain, “Subagents: Multi-agent patterns,” accessed 2026-10-07. |
Code orchestration makes step transitions explicit; it does not remove the need to define what counts as a completed step. OpenAI’s guide discusses orchestration patterns as graph structures, while LangChain shows subagent state examples. Both support composing clear components, not a claim that one hierarchy guarantees reliability.
When should a subtask run synchronously or asynchronously?
Choose based on whether the main answer depends on the subtask’s result and whether the work can happen independently. LangChain’s guidance distinguishes direct synchronous calls from asynchronous start, status, and result operations.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstall- Use a synchronous call when the next step or final answer depends on an ordered result. It is straightforward, but a long-running call can stall the conversation.
- Use asynchronous work when the task is independent or can run in parallel and the user should be able to continue interacting. The workflow needs a way to start the job, check its status, and retrieve its result.
Before choosing, determine what should happen if the work takes too long or fails: whether the user must wait, whether another branch can proceed, and how the system will surface a missing result. Asynchronous execution is not useful unless the parent can later retrieve and handle the outcome.
Best Value
How should state pass between agents and turns?
Separate task instructions from durable conversation state. The child needs the context relevant to its assignment; the overall application needs a deliberate way to preserve information for later steps or turns. Do not assume that a supervisor’s context automatically includes every detail held by a specialist.
OpenAI’s “Running agents” documentation describes several continuation strategies: application-held history, sessions, conversation IDs, and response IDs. Choose one strategy for a conversation unless you deliberately reconcile multiple state layers. Replaying application history while also relying on server-managed state can duplicate context.
For each boundary, identify the minimum information the next owner needs and how it will be transferred. Anthropic’s multiagent documentation, accessed 2026-10-07, labels managed agents beta and describes isolated threads, per-agent configuration, parallelization, and specialization. Those capabilities illustrate why thread and state boundaries need to be designed explicitly; beta status and availability may change.
What hierarchy helps without adding needless layers?
Think in responsibilities, not in the number of agents. A useful architecture assigns ownership clearly, uses specialists only for distinct work, and keeps fixed transitions under application control where appropriate.
- Coordinator: owns the user goal, global state, and final answer.
- Specialists: handle bounded domains with explicit inputs and required outputs.
- Workflow code: controls steps that must run in a fixed order, tracks status, persists state, and handles retries; agents can still make bounded judgment calls within those steps.
- Validator: checks that a required artifact or field exists before the workflow advances. OpenAI’s guide identifies evaluator loops as an orchestration pattern; this is a design choice, not proof of a measured reliability gain.
Every boundary creates another opportunity for context loss or an incomplete response. Add a layer only when it isolates genuinely different expertise, tools, or policies, and define how the parent can tell that the child completed its assignment.
Quick Recap
How do you diagnose an agent workflow that gets stuck?
- It selects the wrong specialist or keeps branching: narrow the allowed roles, make each handoff description specific, and split only when instructions, tools, or policies differ.
- The child worked, but the parent cannot use the result: require the needed findings or artifact in the child’s final output; map critical fields into shared state when a summary alone is insufficient.
- The user experiences a hang: check whether the result is required before responding. Keep dependent, ordered work synchronous; consider asynchronous execution for independent or long-running work, with status and result retrieval.
- A later turn repeats or forgets information: choose where durable state lives and one continuation strategy, or explicitly reconcile the layers you combine.
- The design keeps accumulating agents: improve tool names, parameters, and descriptions first. OpenAI’s recommendation to add agents when tool clarity does not improve performance is a design heuristic, not a universal threshold.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
