Layered by responsibility.
Complexity is earned.
Each layer exists for a specific reason. Select a layer to see its purpose, why it exists and what it depends on.
Execution Strategy Layer
Chooses how much autonomy a task gets.
Complexity is earned: more agents are used only when the task, risk and governance needs justify them.
Model provider abstraction, retrieval
For straightforward or rules-based cases it can be cheaper, faster, easier to audit and safer. A first-class path, not a lower-quality fallback.
Less adaptive reasoning than agentic execution.
One bounded reasoning loop with selective access to approved knowledge tools, when tool choice improves context gathering.
Less separation of responsibilities, but lower coordination overhead, latency and synthesis risk.
Separates evidence, architecture, risk/governance and synthesis when independent specialist analysis is justified; can surface meaningful disagreement.
More latency, cost, failure surfaces and synthesis risk. Not inherently superior — complexity is earned, not assumed.
Separate contracts for structured generation and embeddings, so changing the reasoning provider does not force re-embedding. OpenAI: structured generation + embeddings. Anthropic: structured reasoning/tool selection, not embeddings.
More abstraction code, less provider coupling.
Explicit timeout and retry rules. Permanent validation failures are not retried; timeout, connection, rate-limit and selected 5xx failures may be retried within limits. Safe failure categories are returned instead of raw provider exceptions.