π AgentUse v0.14.0 is here. Our biggest release ever!
It refines the human-in-the-loop & serve experience from 0.13.0 and hardens skills, sandboxing, and auth across the board.
Here's everything that's new π§΅
β οΈ Heads up (potentially breaking): the JSON API moved under an /api/* prefix. The old ?format=json negotiation is removed. Update self-hosted health checks and any scripted JSON consumers.
π Plus lots of hardening: sandbox no longer mounts $HOME, Anthropic OAuth tokens refresh before expiry, stored-XSS fixes on the session page, bounded sandbox exec lifetimes, and more OpenRouter model series.
Full notes π
github.com/agentuse/agentuseβ¦
A $0.05 model nearly beat a $0.95 model on the same task.
Not on speed. Not on cost efficiency.
On pure task completion quality.
I used AgentUse to benchmark 5 AI models (claude opus 4.5, minimax m2.1, glm 4.7, sonnet 4.5, and haiku 4.5) on Notion database CRUD operations. Here's the ranking π
7/ π·οΈ Better Subagent Logging
Sub-agent tool calls are now clearly marked in logs, making it easier to trace execution flow in complex multi-agent systems.