feat(delegation): Advanced Features for Markdown-based Agents

[VascoSch92]

Summary

The PR #2183 added Markdown-Based agents. However, several advanced features are currently unsupported.

This issue tracks the missing features we intend to support in the SDK.

Missing features (ordered by implementation difficulty: low to high)

• max_iteration_per_run: Maximum number of agentic steps before the subagent terminates. (feat(sdk/subagent): Support max_iteration_per_run in file-based agent definitions #2263)
• profile_store_dir: Profile store directory from where loading profiles for model (feat(sdk/subagent): add profile_store_dir #2340 )
• working_dir: working directory for the subagent. [TO DISCUSS]
• condenser: Add a condenser to the subagent. (feat(sdk/subagent): condenser for subagents #2349)
• permission_mode: How the subagent handles permissions. (feat(sdk/subagent): add confirmation policy #2345)
• skills: Skills to load into the subagent’s context. (feat(sdk/subagent): Support skills in file-based Agents #2260)
• hooks: Lifecycle hooks scoped to this agent. (feat(sdk/subagent): hooks for subagents #2347)
• mcp_servers: MCPs available to the agent. (feat(sdk/subagent): add mcp-servers for subagent #2348
• disallowed_tools: Tools to restrict; these will be removed from the inherited or specified list. [TO DISCUSS]
• disallowed_commands: Commands to restrict; these will be removed from the inherited or specified list.
• memory: Persistent memory scope.
• model: Currently, we only support inheriting from the parent model. (feat(sdk/subagent): Support LLM profiles in subagents  #2258)
• background: Always run the agent as a background task. (This is blocked by feat(delegate): Non-blocking background subagent execution #2047 )

Important

Every newly added feature should be backward compatible with the previous markdown format.

[enyst]

model:

It's fine if we don't support more than inherit yet, just for the record: maybe this is a profile name, not a model name exactly. Obviously Anthropic doesn't want more models than their own, but we do. 😹

max_turns: Maximum number of agentic turns before the subagent terminates.

Nit: could we name this max_steps or max_iterations? "Agent turn" means in some contexts (e.g. OpenAI docs and papers) a full agent run until it returns the prompt to the user.

[VascoSch92]

A priori, I wanted to have the same API such that we can also parse a file-based agent coming from Anthrophic. but we can also have an adapter later for that (or just using our parser :-) )

Nit: could we name this max_steps or max_iterations? "Agent turn" means in some contexts (e.g. OpenAI docs and papers) a full agent run until it returns the prompt to the user.

In LocalConversation this corresponds to max_iteration_per_run. Should we name it so?

[enyst]

Sure, sounds good to me, thank you!

(Yes, we do want compatibility, you're right we can add it with an adapter though; funnily enough I was just looking at an OpenAI agent and umm, it seems very simple but not similar? ... because of course 😅 )