Proposal: Add a static-plan execution mode to Bedrock Agents for reproducible workflows

Hi team,

While experimenting with Bedrock Agents, I noticed that the current Agent Planner can produce different execution sequences for the same input, which is expected with LLM-driven reasoning.

But for enterprise workloads (finance, legal, compliance), many teams require:

deterministic behavior

reproducible workflows

auditable step sequences

protection against semantic drift in long tasks

So I built a small drop-in POC that replaces the dynamic planning loop with a YAML-defined static plan + a deterministic step runner.
It does not modify Bedrock itself — only supplements the current planner with an optional “compiled plan” mode.

Repo:
https://github.com/yuer-dsl/bedrock-deterministic-planner-poc

What the POC demonstrates

YAML-based “compiled plan” (static execution graph)

No re-planning after execution begins

Identical behavior on repeated runs

Easy to audit or log for regulated industries

Keeps the original Bedrock flow intact (minimal surface area)

Why this matters

A static-mode option could help Bedrock Agents serve:

finance & banking workflows

insurance & compliance

large enterprise approval chains

reproducible research pipelines

Question

Would the team be open to exploring an optional deterministic execution mode in Bedrock Agents?
Happy to iterate or provide a more detailed minimal POC if helpful.

Thanks for the great work — Bedrock is becoming an important part of many workflows.

Review the linked deterministic-planner POC and compare it with the repository’s Bedrock Agent examples. The issue identifies no repository file or test, so first determine which Agent Planner entry point would be affected and how reproducibility should be verified before implementation.

Written by the indexing model from the issue text.