Tech News
← Home  ·  All topics

Agent Harness

4 GoKawiil briefs on this topic · RSS for agent harness

Memdoor-built multi-agent harness matches GVS5H's Qwen3.8-27B system on LiveCodeBench Hard subset

A reproduction built by Memdoor's coding tool, working only from a written description of the GVS5H paper's method, replicated the five-role multi-agent workflow using Qwen3.8-27B models. Across five passes on the first ten problems from the authors' LiveCodeBench Hard list, the rebuilt workflow averaged 9.8 of 10 correct, versus 9.6 for the authors' own harness and 8.0 for a single Fable 5 model, with full logs and costs published for all 50 runs.

Unreal Labs launches Unreal Agent, an agent harness cutting costs vs Codex

Unreal Labs introduced Unreal Agent, a harness for running AI agents that it says delivers up to 40% cost savings compared to OpenAI's Codex and up to 20% savings compared to Pi, without hurting performance on production workloads and coding/science benchmarks. The company says the harness manages tool calls asynchronously, removing the need for the underlying model to handle waits, polls, and heartbeats itself.

New Show HN site lets developers publicly share their AI tool setups

A developer launched a community platform where engineers can document and share the AI tools, agents, and workflows they use for building software. The site was created out of frustration with scattered, fragmented glimpses of other developers' AI setups seen on social media like Twitter/X. The creator plans to keep his own profile updated with an AI agent and hopes others will contribute their setups too.

Laude releases Headlong, an open-source harness for always-on AI agents

Laude has open-sourced Headlong, an agent framework built around a Bash core under 10,000 lines that keeps AI agents continuously generating thoughts rather than only responding to prompts. Instead of waking on a schedule to run a checklist, a Headlong agent runs a persistent thought stream where incoming messages are treated as observations, and it decides for itself whether and when to reply. Laude tested this internally with a shared agent named Audel across Slack, Telegram, and mobile, letting multiple team members interact with the same continuously running agent.