# Research, alternatives and evidence | HelmLoop

> HelmLoop makes authority, evidence, promotion and observed outcomes explicit across AI agent workflows. Explore the executable 0.1 candidate and the proposed 0.2 protocol.

Source: https://helmloop.dev/research.html

Protocol 0.1 includes eight record schemas, a deterministic six-axis reducer, a CLI and two fixture observation adapters. These are local implementation artifacts, not a public package release or live-runtime certification.

The 0.2 plan adds an assurance axis, authority lifecycle events, criterion-bound outcome observations and versioned evidence admission. These changes are proposed in r3 / sync-4.

Workflows use predefined code paths while agents dynamically choose processes and tools; added autonomy trades predictability, latency and cost.

Modern runtimes already provide graph routing, persistence, interrupts, replay, handoffs, guardrails and traces.

HelmLoop proposes that authority, evidence freshness, promotion and observed outcomes deserve a portable contract boundary when graph completion alone cannot settle a consequential decision.

## Alternative designs

- Keep controls inside the graph: Best when one runtime, one team and one risk model own the full lifecycle. This is the strongest alternative.
- Use a durable workflow engine: Best when replay, queues, timers and long-running reliability are the main problem.
- Use an assurance SOP only: Best when validation and release governance matter but cross-cycle outcome learning does not.
- Use HelmLoop contracts: Potentially useful when several runtimes share one accountable promotion and outcome policy—still unproven across runtimes.

## Primary sources

- [Agent orchestration](https://openai.github.io/openai-agents-python/multi_agent/): Manager-owned specialists and handoffs are distinct orchestration choices.
- [Guardrails — OpenAI Agents SDK](https://openai.github.io/openai-agents-python/guardrails/): Guardrails attach at specific agent and tool boundaries; coverage is not universal.
- [Tracing — OpenAI Agents SDK](https://openai.github.io/openai-agents-python/tracing/): Traces record model, tool, handoff and guardrail events for debugging and monitoring.
- [Building Effective AI Agents](https://www.anthropic.com/engineering/building-effective-agents): Distinguishes workflows from agents and advises adding complexity only when outcomes justify it.
- [LangGraph overview](https://docs.langchain.com/oss/python/langgraph/overview): Describes durable, stateful graph orchestration and human-in-the-loop capabilities.
- [LangGraph interrupts](https://docs.langchain.com/oss/python/langgraph/interrupts): Interrupts persist state and re-execute nodes, making idempotency material.
- [Temporal Workflow](https://docs.temporal.io/workflows): Event history and deterministic replay separate workflow code from external activities.
- [Template agent workflows — ADK](https://adk.dev/agents/workflow-agents/): Documents sequential, loop, parallel and newer graph workflow structures.
- [MCP Architecture](https://modelcontextprotocol.io/specification/2025-06-18/architecture): Places consent, security policy and authorization decisions with the host boundary.
- [MCP Authorization](https://modelcontextprotocol.io/specification/2025-11-25/basic/authorization): Binds access tokens to intended resources and emphasizes least privilege.
- [AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework): Organizes lifecycle risk work through Govern, Map, Measure and Manage.
- [ReAct: Synergizing Reasoning and Acting in Language Models](https://arxiv.org/abs/2210.03629): Shows interleaved reasoning and action; an observation updates a plan but is not a business outcome proof.

## Read the evidence at its actual scope

Current fixture results do not certify a live runtime, broad portability, production readiness or business value. The research page keeps these open questions visible.