Agents that rewrite themselves, and the open models to run them on
Reporting
Self-HPO: a coding agent that tunes its own harness
Researchers from Microsoft Research, the University of Virginia and Carnegie Mellon University describe Self-HPO, a framework in which a coding agent alternates between optimizing its harness and its policy. [1] In each round the current policy analyzes its own failures and proposes harness candidates, then is updated through reinforcement learning under the selected harness. [2] The authors report that Self-HPO raised Qwen3.5-35B-A3B's average resolve rate across three coding benchmarks from 57.8% to 64.8%, and its SWE-Bench Verified resolve rate from 61.0% to 69.5%. [1]
L3M: capability without authority
A paper dated September 30, 2026, with authors listing an Amazon affiliation, presents L3M, an agent runtime written in Lean 4. [3] In L3M all agent state lives in Git, and agents can write sealed Lean programs that become new tools without gaining authority beyond their existing grant. [3] A tool's review tier is derived from whether its effects can be undone or are bounded by a capability token, so human review is reserved for effects that cannot be reversed. [3] On 37 Terminal-Bench tasks, the authors report 41.7% pass@3 for L3M with a shell and 27.0% using only tame tools, against 59.5% for Claude Code with permissions bypassed. [3]
Hy-Memory 2.0 and MemHub
Tencent Hunyuan's Hy-Memory repository describes Hy Memory as a memory engine and Python SDK for self-evolving LLM agents, released as Hy Memory 2.0 (hy-memory 2.0.4) on PyPI. [4] The repository also includes MemHub, described as a local-first desktop and CLI product that unifies conversation history and coding traces from user-authorized AI applications. [4] The README lists an open-sourced Hy-Memory-30B-A3B model on Hugging Face and shows an Apache-2.0 license badge. [4]
Model releases (company announcements)
Meta Muse Spark Open (open weights)
Meta released the weights for Muse Spark Open on October 15, 2026, describing it as a natively multimodal model with 816B total parameters, 42B active, and a 1M-token context window. [5] Meta says the model can run on a desktop using two Apple M5 Ultra Mac Studios with the MLX framework. [5] The weights are released under the Muse Model License with limited commercial terms. [5]
Reflection Beam
A Hugging Face blog post published October 16, 2026 announced Beam, the first model from Reflection, a mixture-of-experts model with 501B parameters and 23B active per token. [6] Beam is text-only, supports a 1M-token context window, is released under the Apache 2.0 license, and is focused on coding and agentic use. [6] The post says the NVFP4 checkpoint shrinks the weights from about 1 TB in BF16 to about 280 GB. [6]
Analysis
Commentary, not verified fact. Two of today's papers take opposite angles on self-improving agents. Self-HPO treats the harness as something the agent should optimize; L3M treats the harness as the thing that must stay trustworthy while the agent extends it. Muse Spark Open is open weights under a limited-commercial licence, not open source; Beam is Apache-2.0 but needs data-centre hardware. Benchmark figures above are self-reported by each team.
Sources
Newsletter
Get QIKA News by email
Short, sourced AI and tech news for builders and small businesses. No spam; unsubscribe anytime.
We use your email only to send QIKA News. See our privacy policy.