SlackAgents: Scalable Collaboration of AI Agents in Workspaces.
| Authors | Zhiwei Liu 0001 et al. |
| Year | 2025 |
| Venue | EMNLP 2025 |
| Paper | View on DBLP |
Abstract
Abstract not yet available in this stub. Read the full paper →
Engineering Breakdown
Plain English
SlackAgents is a framework for enabling multiple AI agents to collaborate and coordinate within workspace environments like Slack. The paper addresses the challenge of scaling agent interactions in real-world communication platforms where agents need to understand context, manage state, and coordinate actions across asynchronous conversations. The system demonstrates how agents can be deployed in Slack workspaces to handle tasks collaboratively, with the architecture designed to handle the complexity of multi-agent communication in a persistent, message-based environment. This work bridges the gap between single-agent LLM deployments and practical multi-agent systems that operate in actual workplace communication tools.
Core Technical Contribution
The core contribution is a scalable architecture for deploying and coordinating multiple AI agents within workspace platforms, specifically designed for the constraints and patterns of asynchronous messaging systems. Unlike traditional multi-agent frameworks that operate in isolated environments, SlackAgents provides mechanisms for agents to discover each other, maintain conversation context, handle concurrent requests, and coordinate actions through natural message-based protocols. The system introduces patterns for agent registration, message routing, context management, and cross-agent communication that are specifically optimized for workspace integrations rather than generic multi-agent orchestration. This represents a shift from theoretical multi-agent coordination to practical, deployable systems that operate within existing enterprise communication infrastructure.
How It Works
SlackAgents operates by registering multiple specialized agents within a shared Slack workspace, where each agent exposes capabilities through command handlers and event listeners. When a user sends a message to the workspace, the system parses the message, maintains conversation history and context, and routes the request to the appropriate agent(s) based on intent or explicit mention. Agents can invoke other agents asynchronously, forming a communication graph where complex tasks are decomposed across specialized AI systems while maintaining the conversation context accessible to all participants. The framework handles state management by storing conversation context in a centralized store, allowing agents to read prior messages and understand the evolving task requirements. Communication between agents happens through message events, with each agent publishing results back to the Slack channel, creating an auditable record of the multi-agent reasoning and decision-making process.
Production Impact
For production systems, SlackAgents eliminates the need to build custom agent orchestration layers on top of workspace platforms—you get native support for multi-agent coordination within tools your organization already uses. This significantly reduces integration complexity compared to connecting external multi-agent systems to Slack via webhooks or custom middleware, as agents operate natively within the workspace context and can leverage Slack's native features like threads, reactions, and file sharing. The main trade-off is that you're constrained by Slack's rate limits and API latency; a single Slack message with multiple agent invocations could add 2-5 seconds of aggregate response time depending on concurrency and downstream LLM inference latency. You'd adopt this when your multi-agent workload is primarily collaborative human-in-the-loop tasks (code review, document analysis, research) rather than high-throughput backend coordination where custom orchestration would be more efficient. The persistent conversation history becomes both a feature (auditability, context reuse) and an operational burden (managing context window growth, storage costs for long-running projects).
Limitations and When Not to Use This
SlackAgents assumes agents operate within a shared, trusted workspace environment—it doesn't provide strong isolation or security boundaries between agents, making it unsuitable for adversarial settings or where you need fine-grained access control. The system's dependency on Slack's API and message queuing means performance is bounded by Slack's infrastructure; you cannot achieve sub-second multi-agent coordination or handle extremely high-concurrency scenarios that would require dedicated message brokers. The paper doesn't address how to handle agent failures, timeouts, or cascading errors in multi-agent chains—there's an implicit assumption that agents are reliable and that error recovery happens through natural conversation rather than formal fallback mechanisms. Additionally, the approach may not scale to very large numbers of agents (10+) operating concurrently in a single workspace without significant message clutter and context confusion, and there's limited guidance on how to structure agent capabilities to avoid redundancy or conflicting responses.
Research Context
This work builds on the growing body of research in multi-agent LLM systems and agent orchestration frameworks (ReAct, AutoGen, LangChain agents), but shifts focus from abstract coordination mechanisms to practical integration with real workplace tools. It addresses a gap in prior research where most multi-agent papers focus on complex reasoning tasks or code execution rather than the pragmatic problem of deploying agents in environments where humans are actively present and expecting natural interactions. The research opens a new direction: treating workspace platforms not as passive integrations but as primary environments for multi-agent systems, which could inspire similar designs for other communication-first platforms (Discord, Teams, Slack competitors). This positions SlackAgents as part of the broader movement toward embodied, tool-integrated AI systems rather than purely abstract agent simulations.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::