Concurrent tools: schedule the dependency graph, not the visible list

Add a dependency edge when a call needs an earlier result, resource version, approval, or committed artifact. Calls that write shared state may need an order even when their argument objects are disjoint. Read-only calls can still require a common snapshot for a meaningful combined answer. Treat shared quotas, provider rate limits, and authorization scope as scheduling constraints. An external service's actual consistency and conditional-write contract determine which apparent operations commute.

Before dispatch, assign stable logical step and effect identities and reserve the resources required by the chosen concurrency policy. A late completion from a replaced worker must not attach to the wrong branch. Preserve output-to-input lineage so the join uses the intended artifact versions. Parallelism should not make a dependent call guess a missing value from a stale cache or model memory. If one branch fails, decide whether independent successful work remains useful or must be compensated under the declared workflow policy.

For fixed task durations, total work is the sum of durations and the dependency critical path is the longest required chain. With P identical worker slots, work divided by P and the critical path are lower bounds on makespan under an idealized schedule. They do not promise that a real queue, network, or external API reaches the bound. Startup, contention, batching, retries, and resource-specific limits can dominate. More concurrency can increase exposed queueing or exhaust a provider quota rather than reduce latency.

Structured concurrency gives each spawned task a scope whose owner waits for or handles its completion. Python 3.14 TaskGroup is one concrete implementation with documented failure and cancellation behavior. It does not make remote side effects transactional or guarantee a child thread or server instantly stops. A durable agent also needs persistent branch identities and joins across worker crashes. Record the accepted completion order or use a deterministic identity-based join when order matters; arrival order is not automatically a stable planning input.

Run the same logical graph with reversed completion order, one delayed branch, duplicate results, a failed prerequisite, and cancellation during a join. Verify no dependent effect starts before its required outcome is accepted, and no result is paired with another branch's arguments. Test shared-budget reservations atomically. Compare serial and bounded-concurrent variants with the same effects and data snapshot. A latency improvement obtained by omitting a required dependency or dropping failed branches is not equivalent execution.

The queue lab isolates one capacity mismatch. Reset, set λ=4 and μ=3, predict q(2)=6, then use four Steps. Separately Reset, set λ=4 and μ=5, predict q(2)=2, then use four Steps. Both start with four abstract work units. The instrument has no dependency graph, worker count, or tool service, and its separate residence-time model has additional assumptions. Use it to see why offered work and service must be measured together, then test the actual scheduler and provider constraints.