If your LLM agent retries a 429
fifty times overnight, retries a 400
three times before giving up, and sends every request to your top-tier model until the run budget is gone, those aren't bugs in the model. They're missing two cheap policies your orchestration layer should be making for you.
AgentFlow4J v0.7.0 (Java multi-agent orchestration on Spring AI) ships those two policies.
A RetryPolicy
that counts attempts is blind to why a call failed. v0.7.0 adds a FailureClassifier
that sorts every failure into one of three categories:
| Category | What the graph does |
|---|---|
TRANSIENT |
|
Retry. If the failure carries a Retry-After hint, that delay is honoured instead of the computed backoff. |
|
PERMANENT |
|
| Stop immediately — no further attempts. | |
OVER_BUDGET |
|
Stop and surface an InterruptRequest so a human can approve more budget and resume. |
The default classifier already understands JDK I/O exceptions, Spring AI / Spring Web 5xx
429
(parsing Retry-After
) vs other 4xx
, and BudgetExceededException
, all detected by class name, so agentflow4j-graph
keeps zero compile-time Spring dependency.
Adding your own rules composes via orElse
:
FailureClassifier domain = cause -> {
if (cause instanceof QuotaExhaustedException) return FailureClassification.overBudget("monthly quota hit");
if (cause instanceof InvalidPromptException) return FailureClassification.permanent("rejected by guardrail");
return null; // unknown — defer to the default
};
RetryPolicy policy = RetryPolicy.exponential(3, Duration.ofSeconds(1))
.withClassifier(domain.orElse(FailureClassifier.defaults()));
Existing policies that only set the legacy retryOn
predicate keep their exact behaviour, the classifier falls back to it when it returns null
.
BudgetPolicy
already caps the cost of a run. v0.7.0 lets it shape which model handles which request:
BudgetPolicy budget = BudgetPolicy.hierarchical(
BudgetLimits.run(5.00), estimator, meter);
// Use "premium" while ≥ $1.00 remains, then "fallback".
RoutingStrategy router = RoutingStrategy.budgetAware(
budget, BudgetPolicy.Scope.RUN, 1.00,
"premium", "fallback");
CoordinatorAgent desk = CoordinatorAgent.builder()
.executor("premium", premiumAgent)
.executor("fallback", fallbackAgent)
.routingStrategy(router)
.build();
While the run budget has more than $1.00
remaining, the coordinator sends work to premium
. The moment less remains, it degrades to fallback
. Reading the live BudgetPolicy.remaining(...)
counter is free, no extra LLM call to "classify complexity."
This is the one cost-aware routing lever that's both deterministic and provably cheaper: classifying complexity ex-ante with an LLM would itself cost a call (chicken-and-egg), and self-confidence routing doubles the cost. Reading counters is free.
The cookbook ships a recipe that wires both features end-to-end:
git clone https://github.com/datallmhub/agentflow4j-cookbook.git
cd agentflow4j-cookbook
mvn -pl 06-cost-aware-routing exec:java
No API key required, Ollama optional, the recipe falls back to a deterministic stub. You should see the first three tickets handled by premium
, then the squad switching cleanly to fallback
once remaining drops below $2.00
. The retry scene then shows a flaky node recovering on attempt 3 with 50ms → 98ms
exponential+jitter backoffs, followed by a classification table.
<dependency>
<groupId>com.github.datallmhub.agentflow4j</groupId>
<artifactId>agentflow4j-starter</artifactId>
<version>v0.7.0</version>
</dependency>
Via JitPack: add https://jitpack.io
to your <repositories>
if you haven't.