[FEATURE] Agent Harness

[mkmeral]

Problem Statement

Strands ships the extension points — hooks, plugins, steering, conversation managers — but developers have to compose them into a working agent configuration themselves. Building a production agent means manually wiring sandbox isolation, context management, permissions, guardrails, and behavioral controls. Every team builds the same scaffolding independently.

The agent harness is the control and composition layer around the agent's decision loop. The model provides intelligence; the harness makes it reliable and directed. We have the primitives but not the composed defaults. We need to close the gap between "here are the building blocks" and "here's a working, safe agent out of the box."

Proposed Solution

Umbrella/epic tracking harness-related work. Two categories:

Missing primitives:

• Intervention primitive — composable control layer for authorization, guardrails, and steering with a shared interface
• Declarative permission boundary — tool-level access control without writing hook code
• Unified memory interface — portable MemoryProvider ABC so backends are swappable
• Agent identity and credential provider interfaces
• Audit trail / compliance logging

Existing primitives that need to become composed defaults:

• Sandbox integration (feat: add Sandbox abstraction for agent code execution environments #1968)
• Token-aware context management ([FEATURE] Automatic Context Optimization strategy proposal #2170)
• Default tools vended in core ([FEATURE] Vend tools from the SDK Directly #2086)
• Execution limits and cost tracking (feat: Add agent execution limits (max_turns, max_token_budget) #2124, [FEATURE] Add LiteLLM Cost Tracking to Agent Results #1216, [FEATURE] Expose total request cost (LLM + tool/KB/DB calls) in AgentResult.metadata — avoid manual per-token pricing #1428)
• Guardrail revamp ([FEATURE] Revamp Guardrail feature #1770, [FEATURE] Provider-Agnostic Application-Level Guardrail (InputFilter & OutputFilter) #1654)
• Lifecycle hook event parity with TypeScript
• Steering expansion ([FEATURE][Steering] Expand steerable events #1386, [FEATURE] Add steer_after_model support to LLMSteeringHandler #1711)
• Agent config graduation ([FEATURE] Graduate Agent Config out of Experimental #1625)

Use Case

Developers building production agents get control and safety out of the box — sandboxed execution, context management, permissions, guardrails — without assembling it from scratch. Every default is customizable through the same extension points we already ship.

Additional Context

Sub-issues are linked below.