𝗖𝗵𝗼𝗿𝗲𝗼𝗴𝗿𝗮𝗽𝗵𝗲𝗱 𝗖𝗹𝗮𝘂𝗱𝗲 𝗗𝘆𝗻𝗮𝗺𝗶𝗰 𝗪𝗼𝗿𝗸𝗳𝗹𝗼𝘄𝘀
Most AI workflows use orchestration. A central controller decides every step. It picks the tool, runs the task, and merges the results. This provides control and a clear path.
My architecture uses choreography instead.
In a choreographed workflow, no single controller rules everything. Instead, each behavior reacts to events. Each component only needs to understand three things:
- The contract it validates.
- The event it follows.
- The event it sends.
This creates a hybrid model. You use orchestration for high-level intent, but use choreography for execution.
Here are the key patterns:
Classify and Act Traditional flows use a router to call a function. Choreographed flows emit events. A message becomes a classified intent. That intent triggers a behavior. Each stage emits a new event. This makes the system modular. You can replace or observe any stage without breaking the flow.
Fanout and Synthesize Instead of one agent splitting tasks, use event subscriptions. One payload can trigger multiple independent behaviors at once. You can have one agent validate types, another check schemas, and a third run benchmarks. Synthesis is not just merging text. It is a semantic reduction to find which path succeeded.
Adversarial Verification Do not trust the first answer. A generator publishes a candidate. Verifiers subscribe to that candidate to attack it. They act as compilers, security testers, or malicious users. If they find a flaw, they emit a counterexample. This counterexample triggers a self-healing pipeline.
Generate and Filter Generation is cheap. Acceptance must be strict. The system produces multiple candidates like fast, safe, or minimal versions. Deterministic gates like unit tests and security policies filter them. Only the winners move forward.
Tournament When there is no obvious best answer, run a tournament. Generate competing candidates and score them against a rubric. Measure CPU cost, memory, and success rates. This turns opinions into measurements.
Loop until Done In most systems, an error ends the flow. In this model, an error is just another event. The system uses the error context to try different healing steps. It tries type conversion, then semantic repair, then user help. This is not a blind retry. It is intent-based healing because the loop has memory.
Koreografi; sistemlerin genişletilmesini, test edilmesini ve iyileştirilmesini kolaylaştırır. Mevcut bir olaya abone olarak yeni bir doğrulayıcı veya yeni bir filtre ekleyebilirsiniz. Graf yapısı bozulmadan kalır.
Kaynak: https://dev.to/fullagenticstack/choreographed-claude-dynamic-workflows-3mgi
İsteğe bağlı öğrenme topluluğu: https://t.me/GyaanSetuAi