👍 Community Priority Poll — What should we build next?

[igordayen]

👍 Vote on what matters to you

Each feature request below has its own comment in this thread. Click 👍 on the comments for the items you want — as many as you like.

Results will be tallied in ~4 weeks and feed directly into the 2.x roadmap.

---

🗺️ Explore the full feature map first

Before voting, browse the interactive feature map — it shows all 113 open discussions organized by category, with keep/triage/close recommendations and descriptions. You can also use it to select your shortlist and copy it as text to paste as a comment here.

---

Categories in this poll

🧠 Agent Architecture & Planning	#1 – #11
🙋 Human-in-the-Loop & Workflow	#12 – #15
⚡ Streaming & Performance	#16 – #20
🔍 Persistence & Observability	#21 – #26
🤖 LLM Providers & Model Catalog	#27 – #34
🔌 MCP & Tool Integration	#35 – #38
🛠️ Developer Experience & Infrastructure	#39 – #43
📦 Data, Modality & Bindings	#44 – #48
📖 Documentation & Examples	#49 – #50

Do not see your idea? Drop it in the comments — if there is enough interest we will add it to the list.

[igordayen]

#1 · 🧠 Multi-agent support / agent orchestration

Orchestrate multiple specialized agents that collaborate on a single goal — delegation, fan-out, result aggregation.

→ #465

[igordayen]

#2 · 🧠 Evolving Mode — event-driven continuous agents

Agents that stay alive and react to external events rather than running to completion on a single trigger.

→ #1725

[igordayen]

#3 · 🧠 Cross-Agent Reasoning propagation

Share reasoning context and intermediate conclusions across agent boundaries so sub-agents benefit from the parent's thinking.

→ #1882

[igordayen]

#4 · 🧠 Plan checker interface

A hook to inspect, validate, or veto a generated plan before execution begins.

→ #1249

[igordayen]

#5 · 🧠 GOAP Stuck State Diagnostic System

Detect when the planner is cycling or cannot make progress, and surface a useful diagnosis rather than timing out silently.

→ #961

[igordayen]

#6 · 🧠 Visualize Agent Planning Graph

Render the planning graph (actions as nodes, precondition/postcondition links as edges) in the developer UI for easy debugging.

→ #1913

[igordayen]

#7 · 🧠 Automatically register actions on domain objects

Allow actions to be declared directly on domain classes rather than requiring a separate agent bean.

→ #1185

[johnsonr]

I think this is potentially too magical. The user can easily add any domain objects using the PromptRunner interface.

[igordayen]

#8 · 🧠 Differentiate entities vs. ephemeral objects on Blackboard

Distinguish long-lived domain entities from short-lived working objects so the planner can reason about identity and lifecycle.

→ #1251

[igordayen]

#9 · 🧠 API for chooseAndRunAgent / chooseAndAccomplishGoal autonomy

A first-class programmatic API for the autonomous goal-selection and agent-dispatch that the shell currently drives.

→ #1248

[igordayen]

#10 · 🧠 Action as an interface

Declare actions via interface rather than annotation so implementations can be swapped or mocked cleanly.

→ #969

[igordayen]

#12 · 🙋 Human Task Framework — inbox, approvals, business workflow

First-class support for pausing an agent to await human approval, form input, or an external review step.

→ #964

[igordayen]

#13 · 🙋 Get more user input mid-flow after agent is triggered

Allow an agent to ask a follow-up question partway through execution without restarting from scratch.

→ #957

[igordayen]

#14 · 🙋 AgentPlatformChatSession — respond to form events

Chat sessions that can receive structured form submissions and button clicks, not just free-text messages.

→ #744

[igordayen]

#15 · 🙋 Debug mode — per-action user confirmation

Step-through debugging where each planned action is shown to the developer for approval before it runs.

→ #661

[igordayen]

#18 · ⚡ LLM response processing speed

Close the performance gap vs. frameworks like crewAI — profiling, parallelism, or reduced overhead per LLM call.

→ #852

[igordayen]

#36 · 🔌 MCP Server: tools not registered without explicit startingInputTypes

Auto-infer startingInputTypes from the tool signature, or emit a clear error when it is omitted.

→ #1183

[igordayen]

#37 · 🔌 Connect Embabel to an existing Spring AI MCP Client

Delegate MCP connectivity to a shared Spring AI MCP client rather than Embabel owning its own connection.

→ #926

[igordayen]

#38 · 🔌 MCP server for Embabel docs

An MCP-accessible documentation server so AI assistants (Copilot, Cursor, etc.) can query Embabel docs directly.

→ #1019

[igordayen]

#39 · 🛠️ Spring Dev Tools support (auto-reload)

Make Spring Dev Tools auto-reload work without breaking the shell and other persistent runners. Most-commented discussion in the repo.

→ #1133

[igordayen]

#40 · 🛠️ Publish to Maven Central

Publish Embabel to Maven Central so enterprise environments that block third-party repos can adopt it.

→ #863

[igordayen]

#50 · 📖 Light theme for the documentation

Add a light theme option to the documentation site — a low-effort improvement with outsized impact on first-time visitors.

→ #1225

[igordayen]

#41 · 🛠️ Kotlin Notebooks for examples

Provide Kotlin Notebook versions of the examples so JVM developers get the same interactive experience as Python/Jupyter users.

→ #967

[neilchaudhuri]

This news may be relevant here.

[igordayen]

#42 · 🛠️ Embabel Evaluation Framework

Structured framework for evaluating non-deterministic agent behavior across runs — planning decisions, action selection, LLM call quality.

→ #1570

[igordayen]

#43 · 🛠️ Framework compatibility — Micronaut / Quarkus

Support running Embabel agents outside Spring Boot — Micronaut and Quarkus as first-class alternatives.

→ #941

[igordayen]

#44 · 📦 Multimodal / media support (files, images to LLM)

Pass images, PDFs, and other media as inputs to LLM calls, not just text.

→ #588

[igordayen]

#45 · 📦 Ability to bind strings from actions

Allow actions to produce plain strings as outputs that downstream actions can consume directly.

→ #1184

[igordayen]

#46 · 📦 Non-Java input types (spreadsheets, Pydantic, prompt templates)

Accept inputs that are not Java/Kotlin classes — CSV rows, Pydantic schemas, raw prompt templates.

→ #611

[igordayen]

#47 · 📦 Integrate with Deep Java Library (DJL)

Use DJL for on-device model inference — embeddings, classification, image models — without a remote API call.

→ #972

[neilchaudhuri]

In general I think some kind of ML pipeline and model integration would be useful and I find SMILE good too.

[igordayen]

#48 · 📦 Tool backoff return

Allow a tool to signal that the agent should back off and retry later rather than treating the call as a hard failure.

→ #1250

[igordayen]

#49 · 📖 StarNewsFinder example — agent naming vs. responsibility

Refactor the flagship example so it clearly separates agent identity from action responsibility, reducing confusion for new users.

→ #1612