Choose a workflow using one concrete task
Example: prepare a release note from code changes and test results. Start with one agent and add coordination only where an independent subtask or a check justifies it.
Act and observe (ReAct)
Read the diff, call a test tool, inspect its output, then choose the next step. A readable reasoning trace is not guaranteed to reveal the model’s actual cause of a decision.
Plan and execute
List the required checks and their dependencies, then execute them. Revise the plan when a test fails; a plan does not grant permission to publish.
Parallel work
Review documentation and run independent test suites concurrently. Give each worker a defined scope and combine actual evidence before writing the release note.
Generate and check
Draft the release note, then check each claim against the diff and test results. Another model’s agreement is weaker evidence than the underlying artifact.
Workflow Visualizer
Scripted dependency diagram: node timing is illustrative. Different shapes show ordering, parallel work and handoffs, not measured speed or reliability.
Sequential
One step at a time
Each node depends on the previous result. This maximizes control and traceability at the cost of speed.
State Management
Persist task status, artifact references and checkpoints with provenance. Define who owns each output and how retries avoid duplicate side effects. Share only the context a worker needs.
Checkpointing
Save state at key points to enable recovery from failures.
Rollback
A checkpoint can restore application state. External side effects need explicit compensating actions; restoring a checkpoint does not unsend an email or undo a purchase.
Handoff Mechanisms
How agents transfer control to each other. Popularized by OpenAI Agents SDK, handoffs enable seamless transitions between specialized agents while maintaining conversation context.
Explicit Handoff
Agent directly calls transfer function with target agent and context.
Condition-Based
Automatic transfer when certain conditions are met (e.g., topic detection).
Escalation
Agent transfers to more capable agent when task exceeds its scope.
Persistent work needs explicit application state
A scheduled run loads a saved goal, relevant memory and the latest external state. It records what changed and stops on completion, budget exhaustion or a missing dependency. Use idempotent actions and explicit authority for publishing or spending. A scheduler does not make a model automatically improve over time.
Enforce permissions in the tool runtime: allowlisted resources and operations, least-privilege credentials, argument validation and approval for consequential actions. UNTRUSTED markers help describe provenance but do not create a security boundary.
For each model call, count uncached input tokens, cached input tokens and output tokens at the applicable rates, then add tool costs. Sequential latency includes each model and tool wait; parallel branches contribute their critical path. Five steps do not imply five times the cost or latency.