Internet of Agents: Weaving a Web of Heterogeneous Agents for Collaborative Intelligence¶
🕒 Published (v1): 2025-01-01 · Source: ICLR · Venue: ICLR 2025 · link
Ask a follow-up
Open an assistant pre-loaded with this paper's context.
💬 Ask ChatGPT✦ Ask Claude
TL;DR¶
IoA is a distributed multi-agent framework modeled on the Internet: heterogeneous third-party agents register on a central server, dynamically form nested sub-teams via instant-messaging-style group chats, and coordinate via a finite-state machine governing conversation flow. With only basic ReAct agents, IoA reaches 40.0% overall on the GAIA benchmark, surpassing prior MAS baselines, and wins 66–76% of head-to-head comparisons against AutoGPT and Open Interpreter on open-ended tasks.
Problem¶
Existing multi-agent frameworks suffer from three compounding limitations: (1) ecosystem isolation — agents outside the framework's own ecosystem cannot be integrated; (2) single-device simulation — real distributed deployments are not supported; (3) rigid pipelines — communication topology, team membership, and state transitions are hard-coded rather than dynamically negotiated at runtime.
Method¶
IoA introduces a three-layer client–server architecture (Interaction / Data / Foundation) with four interlocking mechanisms:
- Agent Registration & Discovery — agents register capability descriptions \(d_i\) on a central server; any agent can call
search_client(L_d) → P(C)to find collaborators matching desired characteristics \(L_d\) via semantic matching. - Autonomous Nested Team Formation — a root group chat \(g_0\) spawns sub-group chats \(g_l\) recursively as sub-tasks require additional expertise, yielding a tree \(h: G \to \mathcal{P}(G)\). Nesting reduces communication channels: \(c_\text{nested}(g) \leq c_\text{full}(g) = \frac{|g|(|g|-1)}{2}\).
- Finite-State Machine Conversation Control — group chat state \(M = (S, \Sigma, \delta, s_0, F)\) with states \(S = \{s_d, s_s, s_a, s_p, s_c\}\) (discussion, sync/async task assignment, pause-and-trigger, conclusion). Each agent's LLM autonomously computes \((s_{t+1}, c_{t+1}) = f_\text{LLM}(M_t, s_t)\) to select next state and next speaker.
- Task Assignment & Execution — tasks \(t = (d_t, S_t)\) are allocated synchronously (chat paused until completion) or asynchronously (parallel execution), with an
Agent Integration Blockproviding a uniformrun: String → TaskIDinterface for third-party agents.
Key Contributions¶
- An agent integration protocol enabling heterogeneous third-party agents (AutoGPT, Open Interpreter, etc.) running on different devices to join a shared collaboration space.
- An instant-messaging-inspired architecture with server-side agent registry, group-chat routing, and client-side wrappers.
- Autonomous nested team formation that reduces communication complexity compared to flat fully-connected topologies.
- A Speech-Act-Theory-grounded FSM for group-chat flow control with LLM-driven state transitions and speaker selection.
- Demonstrated cross-domain applicability without task-specific prompt tuning across GAIA, open-ended instruction, embodied AI, and RAG benchmarks.
Results¶
- GAIA benchmark (validation set): IoA (4 basic ReAct agents) achieves 40.0% overall vs. AutoGen 39.39%, FRIDAY 34.55%, GPT-4+Plugins 14.60%, AutoGPT-4 4.85%.
- Open-ended instruction benchmark (153 tasks, 4 categories): IoA wins 66.7% vs. AutoGPT and 76.5% vs. Open Interpreter overall; strongest in Math (83.3% vs. both) and Coding (83.3% vs. AutoGPT).
- GPT-3.5-based IoA on RAG QA achieves performance close to or exceeding GPT-4 baselines and surpasses prior MAS on the same task.
Limitations¶
- Central server is a single point of coordination; scalability and fault tolerance under large agent counts are not evaluated.
- Semantic matching for agent discovery quality and latency are not ablated.
- FSM state transitions depend entirely on LLM judgment; failure modes (infinite loops, stuck states) are not quantified.
- Experiments use GPT-4-1106-preview; generalization to smaller or open-weight LLMs is not assessed in the main paper.
- Cost analysis and nested-team formation precision are deferred to appendices, not the main results.
Relevance to Harnesses / Meta-Harnesses¶
IoA is a meta-harness in the precise sense: it provides a protocol layer and orchestration runtime that wraps arbitrary third-party agents without modifying them, coordinates their discovery and teaming dynamically, and governs execution flow via an FSM — all concerns that define a meta-harness rather than a task-specific agent. The agent registration/discovery mechanism directly instantiates the "capability registry" pattern central to composable harness design. The nested team formation with recursive sub-group spawning mirrors the hierarchical task decomposition seen in agentic harnesses like AutoGen, but externalizes the topology to runtime negotiation rather than compile-time wiring, which is a key design axis for harness extensibility.