Future ideas: True REPL, First-class multimodal generation, more open-source models, internal MBUS, notebook client, massive parallelism

[iqdoctor]

Future ideas (as brainstorming items):

• True REPL — a Read–Eval–Print Loop inspired by Lisp and Mathematica. Update any prompt, orchestration script, or model on the fly, even during execution. This feature could dramatically speed up development and let multi-agent systems evolve much faster.

• First-class multimodal generation. Right now MCP has to be used to generate images; ideally, multimodal outputs should be native results of prompts.

• Extended model list, including exotic providers like GigaChat — for example, LangChain already supports it. Note: GigaChat available as open source platform. @Amigol show me some Cursor gentrated code for this.

• A small internal MBUS — an event/message bus with a publish/subscribe model, allowing agents to be event-activated and fully asynchronous, with an external API.

• A Mathematica-style notebook client.

So the direction is: massive parallelism, event-driven activation, openness to novel models, fully asynchronous behavior, and interactive development throughout.

• And of course, finish point 0 for me — the agents_as_tools massive-parallelism feature:
• • Feature: Agents-as-Tools — expose child agents as callable tools with parallel execution #458
• • Feat: hybrid Agents-as-Tools/MCP-as-tools experimental agent #515

[evalstate]

Yes to all of this :)

True REPL — a Read–Eval–Print Loop inspired by Lisp and Mathematica. Update any prompt, orchestration script, or model on the fly, even during execution. This feature could dramatically speed up development and let multi-agent systems evolve much faster.

It's largely refactoring to do this now; some earlier design decisions now reversed were prohibitive of this. Model switching should now be easy, as well as MCP Server attachment/detachment and in-band context editing. My main thought here is strategically managing the effort. Things exposed as / commands that ACP Client can use are great. Things like context/history editing (like double ESC navigation) which are likely to end up in ACP are probably worth supporting through that channel.

First-class multimodal generation. Right now MCP has to be used to generate images; ideally, multimodal outputs should be native results of prompts.

The current design of fast-agent should make this very straightforward, as the PromptMessageExtended format makes this easy for inputs of docs/images etc:

For native image outputs, I'd imagine we add it as a ContentBlock to the result.

Extended model list, including exotic providers like GigaChat — for example, LangChain already supports it. Note: GigaChat available as open source platform. @Amigol show me some Cursor gentrated code for this.

Main point here is that one of the things sets fast-agent apart is that listed models are comprehensively tested e2e (for multimodal, structured, tool calling, message sequencing etc.). It may be that we draw a visible line somewhere (primarily for support reasons; e.g. 8b models are very useful in fast-agent, but they're not going to make good orchestrators.

A small internal MBUS — an event/message bus with a publish/subscribe model, allowing agents to be event-activated and fully asynchronous, with an external API.

Makes sense, and would make them easier to wire up. I would dearly love to have A2A support and this sounds like a good enabler.

👍