This week in deep learning, we bring you Haiku 5.5, Learning Harness Engineering by Building a Coding Agent From Scratch and a paper on Self-Compacting Language Model Agents.
You may also enjoy Meta’s Muse Wonder, Redesigning source control at Anthropic, a paper on Recursive Self-Improvement through Multi-Agent Self-Supervision, and more!
As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.
Until next week!
Industry
Haiku 5.5 drops to 0.10/0.50 per million tokens below 100,000 context and scores 39.2% on Terminal-Bench 4.0 at max effort against GPT-6 Luna’s 16.4%.
Muse Wonder Full scores 56 on the Artificial Analysis Intelligence Index but ships subscription-only, while the API-available Wonder reaches 52 at 4/20 per million tokens.
Jev creator TypeSafe closes $870M round at $7.5B valuation
TypeSafe raises $870 million at a $7.5 billion valuation led by a16z with Sequoia and DCVC, under a month after its Jev decision model launched.
Introducing Clef-omni with full multimodality, plus a faster Clef and a cheaper Clef-flash
Cloudflare extends its Clef decision models to native audio, video and image input on a Qwen3-Omni-30B-A3B base, returning text decisions in about 130ms at the median.
MLOps / LLMOps / AgentOps
Run multi-node inference workloads using Ray LLM Deep Learning Containers
A hands-on guide about splitting one LLM across two GPU nodes on EKS with separate prefill and decode workers behind a single OpenAI-compatible endpoint.
Building agents that transact with Anchor Browser and AgentCore payments
A guide about letting browser agents pay for x402-protected resources while AgentCore checks budgets and signs transactions without exposing wallet credentials to the agent runtime.
Learning
Redesigning source control at Anthropic
An unusually candid engineering article about replacing git with an in-house system in five months with seven people, after merges grew from 2,500 to over 30,000 per day.
Investigating unintended model actions in our evaluations and internal use
A transparency report about four categories of Claude overstepping — exploiting a server flaw, submitting a real form, bypassing paywalls, evading fetch limits — prompting Anthropic to cut all internal evals off the live internet.
A methodologically careful report about grading six frontier models on real tasks drawn from Epoch’s own work, finding them reliable on well-defined tasks and failing on the open-ended parts.
Learning Harness Engineering by Building a Coding Agent From Scratch
A webinar write-up about harness engineering from first principles, distilling an eight-article, six-video open-source course on building a coding agent into 90 minutes.
Libraries & Code
An open-source AI observability tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.
K-Dense-AI/scientific-agent-skills
Turn any AI agent into an AI Scientist.
Papers & Publications
Self-Compacting Language Model Agents
Long agent traces composed of chains of thought and tool calls accumulate stale content that anchor subsequent generations, and eventually outgrow the context window. Existing scaffolds mitigate it with fixed-interval compaction triggered at a token threshold. Such triggers pay no heed to trajectory structure, risking discard of partial results mid-derivation or mid-search. We propose SELFCOMPACT, a scaffold that allows the model itself to decide when and how to compact. Specifically, it pairs two inference-time elements: (i) a compaction tool the model invokes to summarize the accumulated context, and (ii) a lightweight rubric specifying when to fire (a sub-task has resolved, or the trajectory is converging) and when to suppress (mid-derivation, or when stuck). Both are needed. The tool alone is unevenly used across open-weight models, often invoked at unhelpful moments or not at all; the rubric alone cannot act. Together, they elicit effective adaptive compaction without any fine-tuning or external supervision. We present empirical results on six benchmarks (competitive math and agentic search) and seven models. Our results show that SELFCOMPACT matches or exceeds fixed-interval summarization at a fraction of the token cost, improving over a no-summarization baseline by up to 18.1 points on math and 5–9 points on agentic search at 30–70% lower per-question cost. Our results expose a meta-cognitive gap: although unprompted models cannot reliably tell when their own context is rotting, a lightweight rubric closes this gap, reframing when to compact as a capability that scaffolds can supply without training.
Recursive Self-Improvement through Multi-Agent Self-Supervision
Recursive self-improvement (RSI) of a model on non-verifiable tasks, such as open-ended research, faces a supervision bottleneck when its outputs exceed what even human experts can reliably assess, leaving the model itself (optimizee) as the best available optimizer and evaluator. However, a single model instance struggles to critique and improve its own complex reasoning under this homogeneous loop. To address this, we propose Multi-Agent Self-Supervision (MASS), an RSI method that alternates between evolutionary workflow optimization and supervised fine-tuning on self-generated trajectories. Guided by early findings that multi-agent topologies excel at complex reasoning, MASS prompts a single base model to iteratively propose, execute, and self-evaluate multi-agent workflows. Through an evolutionary search constrained by structural guardrails, the model optimizes these computational-graph-like orchestrations, discovering the most effective distinct roles and information routing for a given task. Over two MASS cycles with Qwen3.6-27B, the model achieves 1.2-1.6x higher performance per output tokens on four open-ended public benchmarks. Because the improved model subsequently acts as a better optimizer and evaluator, this alternating framework enables a continuous, recursive bootstrapping of the model’s capabilities. Moreover, multi-agent traces are also more training-efficient: a student trained on them outperforms a single-agent student trained on 1.4x more training tokens. These findings suggest that jointly learning orchestration and bounded subagent execution from multi-agent trajectories can provide an effective signal for RSI.
