More agents cannot shorten a dependency chain.
Use multiple agents when work can be split into independent outputs with clear contracts and a reliable integration step. Use one bounded workflow when the tasks share state or require a sequence. Parallelism can reduce elapsed time, but it adds coordination, cost and conflicting-write risk. There is no supported universal 70% speedup in this guide.
What changed—and what the sources establish
LangGraph documents persisted workflow state through checkpointers. That supports resumption and human review, but application side effects still need safe replay and reconciliation. LangGraph persistence.
Microsoft describes Agent Framework as the successor bringing together AutoGen and Semantic Kernel concepts. For a new project, evaluate that current framework and its migration documentation rather than assuming an old AutoGen tutorial is the default architecture. Microsoft Agent Framework overview.
CrewAI documents sequential and hierarchical processes. Those are orchestration choices; naming several roles does not by itself create independent work or prove a productivity improvement. CrewAI processes.
Choose the approach that fits the constraint
Draw the task dependencies and state ownership before choosing a framework.
| Situation | Start here | Tradeoff / acceptance check |
|---|---|---|
| Independent research slices | Parallel agents with evidence contracts | Integration checks overlap, sources and contradiction |
| Dependent action sequence | Sequential durable workflow | Persist state; do not parallelize required prerequisites |
| High-risk proposed change | Generator plus bounded independent review | Require evidence, avoid endless reviewer loops |
| Shared mutable business record | Single writer and serialized commands | Prevent conflicting actions and stale-state writes |
Split on ownership and dependencies
List the outputs a customer actually needs. For a vendor evaluation, independent agents might inspect security documentation, pricing terms and integration requirements. Give each a defined scope and a structured evidence format. The integration step should identify contradictions and missing information rather than merely concatenate three reports. If all agents are browsing the same pages and writing the same conclusion, parallelism mostly buys duplication.
Shared files and live records require stronger coordination. Assign one owner to each output or serialize mutations through a command service. An agent updating a customer’s address while another creates a shipping label can invalidate the second action. Re-read authoritative state at execution time and keep business permissions outside agent discretion. A communication channel is not a transaction boundary.
Use durable workflow state with bounded retries
Model the workflow as explicit states: pending evidence, ready for review, approved for action, action submitted and reconciled. Store enough information to resume after a process restart. A checkpoint should reference tool evidence and command identifiers, not only a natural-language narrative. Limit retries and route exhausted or ambiguous cases to an operator with the attempted steps and current state.
Put external side effects behind idempotent or deduplicated commands. A resumed agent may repeat a reasoning step, but it must not repeat a payment, email or inventory mutation accidentally. Distinguish retryable reads from writes with uncertain outcomes. Record each attempt’s cause and outcome so another worker can recover without guessing what the previous model probably intended.
Choose the framework after the control requirements
Start from requirements: persistence, human review, tool isolation, typed outputs, observability and existing language/runtime. A graph-oriented design suits explicit dependencies; a hierarchical design can suit delegated subprojects with a manager and bounded outputs. Verify current framework documentation and support status. The easiest demonstration is not always the easiest system to operate when an agent stalls halfway through a customer action.
Keep model and framework comparisons separate where possible. A workflow can improve because the task contract becomes clearer even if the model stays the same. Instrument time spent executing, waiting, reviewing and integrating. Track total token/tool cost and accepted outputs. A faster research draft that consumes more reviewer effort may fail the objective despite a shorter wall-clock demo.
Worked example: illustrative assumptions, not a reported result
Illustrative critical path: three independent analyses take 20, 30 and 40 minutes. Sequential completion takes 90 minutes before integration. Parallel completion takes 40 minutes, plus an assumed 15-minute integration step, totaling 55 minutes. The hypothetical reduction is 35 ÷ 90 ≈ 38.9%, not an automatic 70%.
If each analysis depends on the preceding output, the dependency chain still takes 90 minutes; agents cannot wish it away. If integration takes 50 minutes because outputs conflict, parallel completion reaches 90 minutes anyway. These invented numbers show what to measure: independent work, critical path and integration overhead. Cost and review quality remain separate acceptance conditions.
Failure modes to catch before scaling
- Duplicate delegation: several agents solve the same task without a purpose. Assign exclusive scope or an explicitly bounded review role.
- Shared-write races: concurrent agents act on stale state. Use one writer and revalidate state before commands.
- Coordination loops: agents keep requesting revisions without a stopping rule. Set output contracts, budgets and an escalation owner.
Parallel agents are useful when the dependency structure permits them. A failure of task design should be fixed at that boundary, not hidden by adding another manager or reviewer.
A concrete next step
Draw the dependency graph for one recurring workflow. Parallelize only independent nodes, assign one integration owner and demonstrate restart plus duplicate-command recovery before expanding the agent count.
| Contract | Each task has inputs, output schema and acceptance |
|---|---|
| State | Named writer and authoritative record for each mutation |
| Budget | Time, tool and iteration limits per workflow |
| Evidence | Critical path, integration effort and accepted outcome |
Compare the proposed workflow with a simple baseline using the same objective. Report elapsed completion time, active human effort, cost and failure recovery. If a framework upgrade changes checkpoint or tool behavior, repeat the recovery checks before changing production execution.
Practical questions
Do multiple agents always make a workflow faster?
No. Parallelism helps independent work. Required dependencies remain sequential, and integration or review overhead can remove the apparent speed benefit.
Which framework should a new team choose?
Choose after specifying persistence, review, state ownership, runtime and observability requirements. Compare current LangGraph, Microsoft Agent Framework or CrewAI documentation against those requirements.
Can agents safely edit the same business record?
Use a single writer or serialized command service with authoritative-state checks. Natural-language coordination alone does not prevent stale reads or conflicting mutations.
What should replace a promised 70% speedup?
Measure the actual critical path, integration effort, accepted output, total cost and human review against a comparable baseline. Report the conditions and uncertainty.
Primary sources checked 11 October 2026
- LangGraph persistence — Checkpointers save workflow state for resumption and human-in-the-loop operation.
- Microsoft Agent Framework overview — Microsoft describes Agent Framework as the direct successor combining AutoGen and Semantic Kernel concepts.
- CrewAI processes — CrewAI documents sequential and hierarchical task processes.
The calculations in this guide are illustrative planning examples. Product documentation establishes supported features and conditions; it does not establish a measured return for your organization. Recheck account availability and release notes when implementing the recommendation.
Continue with Agent control plane and Agent workflow evidence.