Summary
Agent-E (Emergence AI) is a text-only web agent with a two-level hierarchy: a planner agent decomposes the user task into sub-tasks and a browser-navigation agent executes each one with a small set of primitive skills (click, type, get DOM, open URL, press keys). The navigation agent chooses among three DOM distillations (text-only, input fields, all fields) per step, and after every action receives a “change observation” describing what changed on the page. On the WebVoyager benchmark it reaches 73.2% task success, about 20 points over the previous text-only SOTA and 16 over the multimodal SOTA, and it is the first to report task time, LLM-call count and error-awareness alongside success rate. The paper distils the work into design principles for agentic systems generally.
Key points
- Hierarchical architecture: planner (task management, verification, backtracking) insulated from DOM noise; navigator freed from global planning.
- Flexible DOM distillation: three denoised representations of the accessibility tree/DOM, selected per sub-task; denoising is what keeps long tasks inside the context window.
- Change observation: verbal feedback on the state delta after each action (cf. Reflexion), improving state awareness.
- Results: WebVoyager 73.2% success overall, up to +30% on some sites (e.g., Wolfram Alpha); metrics beyond success: completion time, number of LLM calls, error awareness (does the agent know it failed?).
- Human-in-the-loop mode for logins, CAPTCHAs and ambiguous tasks; noted that no benchmark measures HITL agents.
- Design principles: a well-chosen set of primitive skills; hierarchy only when tasks decompose; denoise observations; give linguistic feedback of actions; support human-in-the-loop; agentic self-improvement from experience.
Relevance
- Web Agents — the agent-side paper in the list: shows which architectural choices move web-agent success, and argues success rate alone is a misleading metric (retries, cost, time, error awareness matter). Complements the benchmark papers WebArena, WorkArena, WorkArena++.
Notes
(batch ingest; no notes yet)