[FEATURE] Load Agentic Configurations from a Dictionary

[vawsgit]

Problem Statement

I would like strands to allow the serialization of agent and tool configurations to/from a dictionary structure so that I can persistence/load agents and tools from a configuration store like a database or file without requiring code to be updated for configuration changes like tools, prompts, and other non-code elements of the agentic configuration.

Proposed Solution

Provide a strands specific DSL for describing agent and tool configurations.

Use Case

The GenAIIC IDP Accelerator would like to make agentic intelligent document processing workflows configurable by the user without code. If it were possible to load an agentic configuration from a dictionary, the system could ship with an agentic configuration that could be updated by the user and adapted for their workload.

Alternatives Solutions

No response

Additional Context

No response

[kaleko]

+1 to this. This would also enable agents to easily build other agents, which is an exciting direction.

[zastrowm]

This is something we're interested in looking at, but need to see some design/specifications before jumping to implementation - POCs can be used to tease out details though.

Some notes/call-outs:

• Code stuff:
  • Is this a one way or a two way process - can you round-trip the configuration?
    • If round-tripping is implemented, does it intersect with [FEATURE] Add reset() method to Agent class for clearing all agent state between runs #329?
  • Is this built into the SDK Agent or is this built "on-top" of our extensibility mechanisms?
  • What properties should be configurable and what shouldn't be? What happens when a given code dependency (like a specific tool) is missing in code but specified in the configuration
• File format stuff:
  • Are there any existing standards that we can rely on for configuring clients?
    • aws/amazon-q-developer-cli/blob/main/docs/agent-format.md is one such format that we're interested in exploring
  • When we have other languages, does the config format carry over to other languages?
  • Do we need a way to save a running agent's setup to a file?

If anyone is interested in working through some of these questions, starting a GitHub discussion would be a great place to do that.

[vawsgit]

Thanks for taking a look, I started a discussion HERE
You can take a peek at the proposed implementation HERE

The main goal was to wrap the existing Strands Python API around agents, tools, graphs, and swarms with a DSL configuration. Wherever possible I tried to keep with existing Strands conventions and naming.

Code Stuff:

• This is a 2x way process there is a serialize_agent() method here that will return an Agent() as a Dict. And similar methods for graph and swarm
  • That said, the use case where an agent configuration would be persisted from a code first agent, and then reloaded during the same runtime process such that it would require a reset is not prohibited, it's just not a use case I am focusing on so I didn't implement a reset with this first experimental version. It may or may not require modification to function. I would offer that can be left for a future version. The functionality being proposed would be consumed by systems like the GenAIIC IDP Accelerator, that was the use case the PR addresses.
• This is not built "on-top" but for now I set it aside as experimental and if anyone wants to use it, they can call the AgentConfigLoader() directly. So the proposed implementation touches no existing code, it's completely contained in the strands.experimental.config_loader module. See the FUTURE proposal for modifying Agent(), Graph(), and Swarm() constructors here.
• Many reasonable properties are configurable. If a code reference is provided that does not exist or can't be referenced in scope an exception would be raised. The configuration loader supports many complex scenarios, but is not meant to fully replace Python coded agents. There may be features supported by the Python API that aren't/can't be supported by the configuration loader.

File format stuff:

• For the configuration file format, I chose to follow the existing Strands Python API, tools:, agent:, graph:, swarm:, etc. I think it came out really well, and there are pretty much 1:1 correlations between the Python API and the dictionary configuration that should be readily apparent and easy to follow for any Strands developer. I would not recommend trying to use the configuration loader as an adapter to any common format that the Strands Python API does not already follow.
• Do note the schema provided that can be used to validate YAML configurations in IDE tools like VSCode. These schemas work well for having a GenAI coding agent generate an Agent configuration, and could be helpful for Strands MCP tools that could create Agent configurations.
• The entire premise of the implementation is for cross platform interoperability between Strands and non-Python persistent stores (JSON, YAML, DyanamoDB, Postgres, etc.). By design the configurations must carry over to other languages.
• Saving an Agent configuration to file is supported by the serialize_ methods. It's not a primary use case for our customers so it was not a focus of the impending PR.

Thanks for taking a look at this feature.