DailyWeights & Takes

2 Nov 2026 · agents on a budget

The last 24 hours, to 2 Nov 2026 10:16 ET

Edition 14 · 2 news · no releases · 2 repos · 5 papers

News 2

1. SemiAnalysis2 Nov

   GDPx: on SemiAnalysis's long professional tasks, the best agent scores 12%

   Eleven equally weighted domains, each built with a specialist partner: chip design, cybersecurity, biology, finance, legal, quant trading, hardware, materials science, alignment, software engineering and "recursive self-improvement". Tasks run up to 8 hours in simulated environments, using each model's own harness where it has one. Claude Opus 5.5 leads at 12.0, ahead of Gemini 4 Carbon (10.9) and GPT-6 Astra (9.6). Binary pass rates in finance and legal are close to zero for every model, so the authors also report partial credit (Opus reaches 44% on a legal task). They also say the safety classifiers on Opus 5.5 and GPT-6 Astra pulled down their cyber, bio and alignment scores. Many tasks are private and no harness is published. Agents still fail about nine in ten long professional jobs, and in some domains the provider's safety filter, not the model, sets the ceiling.

   13 models rankedpartly privateleaderboard

2. CNBC2 Nov

   Anthropic's amended S-1 adds the third quarter: about $17 billion in revenue and a $23 billion net loss

   Revenue for the first nine months was $33.2 billion, about 13 times a year earlier. Most of the $104 billion nine-month loss comes from fair-value changes on convertible notes, not operations. The line that matters for builders is purchase commitments and leases of about $895 billion, which the filing says are "substantially all" compute and infrastructure, while compute spending rose 329% over the nine months. Plan allowances are priced against a compute bill that is growing faster than revenue, so expect them to keep changing, and budget each task in tokens rather than assuming a flat plan covers it.

   31 ptsfilingdiscussion

A short list for a Monday morning. We skipped an alleged leak of an Anthropic checkpoint on Hugging Face: nothing backs it up, and commenters who inspected the files found mostly noise.

GitHub repos 2

1. GitHub1 Nov

   Balanov/uncut

   A headless coding-agent harness in one 7 MB Go binary with four tools: read, write, exact replace and run command. It sends the whole conversation on every call, never caps file or command output, and exits with an error when the provider truncates a reply, instead of passing off half an answer as done. Works with any OpenAI-compatible endpoint, including Ollama. Not a sandbox. MIT.

   We're skepticalWith a large context window, a full harness like Claude Code or Codex rarely cuts anything that matters, and it adds caching, permissions and a sandbox. This mostly suits scripts and small local models.

   554 ★new, created 1 Nov

2. GitHub25 Oct

   emanuele-em/kanzo

   A self-hosted kanban board that coding agents use over MCP: they take cards, move them and comment, and find the project from the repo's git remote. One 4.5 MB Rust binary with SQLite inside. Its tool list costs about 900 tokens once per session and replies run 10-60 tokens, by the author's own measurement. Share it live over a LAN or VPN, or sync through a private git repo. AGPL-3.0.

   +330 ★ in window≈365/day · was ~66/day (breakout)

Two this time. The rest of the window's breakouts already ran this week, and most of the other new repos were near-identical skill packs.

Papers 5

1. Hugging Face2 Nov

   R²Orch: an agent orchestrator that trains away its own skills

   In an inner self-play loop, the orchestrator improves on tasks a proposer picks to keep outcomes varied. In an outer loop, each skill is withdrawn once a controlled test shows the orchestrator does as well without it. Failures become skills, skills become weights, and the target is deployment with no skills at all. The authors report gains over baselines on twelve datasets covering QA, math, code and interactive tasks. Today a skill sits in your prompt and costs tokens on every call; here skills are scaffolding the model learns and then drops.

   68 upvotescode

2. Hugging Face2 Nov

   Let a cheap predictor decide what the expensive model runs

   A theory paper on a predict, gate, execute, update loop. A light "System-1" scores each candidate action (a rollout, a reasoning branch, a tool call) and only those above an adaptive threshold reach the expensive "System-2". With perfect screening, execution compute drops toward the share of candidates that turn out useful. Prediction error, exploration and distribution shift eat into that, and the authors bound how much. A formal case for putting a small model in front of your agent's tool calls.

   50 upvotestheory

3. Hugging Face2 Nov

   HALO: proactive agents fail mostly by stepping in too early or doing too much

   400 scenarios across 10 everyday domains, where the agent has to find the evidence and then decide whether, when and how far to help without being asked. 20 experts annotated 8,000 responses from 10 models. The best proprietary model was acceptable 72.9% of the time and the best open model 56.6%. Premature timing or excessive intensity accounted for 61.2% of rejections. As always-on assistants spread, restraint is the skill to test before you hand one your inbox.

   6 upvotesdataset

4. Hugging Face2 Nov

   CodeSpecWeave: coding tasks whose requirements constrain each other

   One agent grows real developer tasks into coupled requirements, such as interface compatibility and error handling, with executable tests. A second agent audits solutions from several models so the tests neither miss flaws nor reject valid fixes. The result is 4,670 training tasks from 2,865 repositories, and RL on them beats existing baselines on DeepSWE-v1.1. A 100-task, 1,044-spec benchmark leaves frontier models clear headroom. Interacting constraints are what break agent PRs in real codebases, and most benchmarks don't test them.

   5 upvotesbenchmark

5. Hugging Face2 Nov

   Why RL fine-tuning drifts off-policy, and a fix for it

   When an inference engine samples and a separate trainer updates, the gap between them has two sources. Engine numerics add a near-constant bias that is easy to cancel. Stale sampler weights create a lag that feeds back into itself. On Qwen3-1.7B, importance weighting broke down at the step where the ratio's effective sample size fell below one half. The proposed correction, DLC, resyncs the sampler at that point and kept small Qwen3 runs close to on-policy. Useful if you run asynchronous RL and have seen it collapse for no obvious reason.

   5 upvotesRL training

Things to try

The take

Get the take on X. Follow @weightsandtakes for one post a day: the take, and a link to the new edition.