The Pulse, Thursday 2 October 2026
Claude Code gets mods—TypeScript plugins that change behavior and UI, built by Claude or handwritten. Cloudflare's Clef decision models launch with an RL fine-tuning platform on Workers. arXiv delivered overnight: voice-agent document grounding, agentic coding under RL, and agent memory.
Models
Apodex 1.1 Mini (free)
Reasoning-first model for long-horizon research and forecasting, works with files, data, code, and tools, 262k context, free.
Free 262k reasoning model: fits Mackinac legislative research before we pay for a bigger one.
Pareto 26.10 Preview
Composite multimodal model: runs several models per request, evaluates answers, returns the best one. 1M context, $0.80 / $3.20 per million tokens.
Multi-model voting behind one call: could replace a DSPy judge ensemble with one model that does it internally.
Research
DuplexSpeechBench-Document Grounding: Benchmarking Document Grounding and Hallucinations in Voice Agents
Benchmark for document grounding in voice agents across five professional domains, tests three failure modes (context saturation, instruction conflict, fabrication).
We grade EverRaise voice agents on grounding accuracy; a specialized benchmark tells us whether the general eval is missing anything.
Cross-Benchmark Transfer from RL on Agentic Coding Tasks
RL on expert-built agentic coding tasks closes the last-mile gap: dropped requirements, incomplete test coverage, broken behavior. Tests cross-benchmark transfer.
Coding agents that test themselves incompletely and break existing behavior are the current bottleneck for autonomous PRs.
MemFit: Efficient Long-Term Agentic Memory
Long-term memory system for conversational agents: bypasses costly LLM-driven memory organization and consolidation, optimizes write operations.
Botterfly's Mnemosyne does exactly this; a paper on memory consolidation costs validates the design target.
Beyond Leaderboards: Tokenomics of Agentic Small Language Model Ensembles
Evaluates SLM-judge-mediated agentic ensembles on operational reliability, cost, latency, and token efficiency, not just accuracy.
Our wehshi PR review panel runs three models with a judge; a tokenomics framework tells us where the cost is and whether cheaper models change it.
LawCompass: Navigating from Legal QA to Multi-Agent Deep Research with Grounded Evidence
Multi-agent system for legal research: systematic evidence retrieval, multi-step reasoning, grounded answers. Moves beyond conversational QA.
Pipeline's nonprofit clients need grounded policy research; a legal multi-agent pattern is one model for similar legislative workflows.
Releases
Langfuse v4.49.0
Org-level analytics view, ingestion overview route, tool filters on session table, evaluator onboarding tracking.
We trace SalesKick and EverRaise flows through Langfuse; org-level ingestion stats and tool filters make debugging multi-service flows faster.
Hacker News
Clef: Open-weight decision models, and new RL fine-tuning platform
Cloudflare launches Clef, open-weight decision models (Yes/No/Confidence), and an RL fine-tuning platform on Workers.
Decision models can gate agent actions live; an RL fine-tuning platform on Workers fits our serverless edge deployment pattern.
RIP, vector database
Turbopuffer argues vector databases are becoming unnecessary: purpose-built compute plus object storage beats them on cost and latency.
We run Pinecone for Mackinac; if the economics flipped we could swap it for cheaper cloud-native storage before we scale further.
GPT-Synopsys: Frontier Intelligence to Revolutionize Chip Design
OpenAI and Synopsys announce GPT-Synopsys, frontier models for chip design workflows.
No direct use for us yet, but another vertical reasoning model worth watching: if chip design can ship one, legal and policy research will too.
People on X
Claude Code mods
Claude Code now supports mods: TypeScript plugins that customize behavior, UI, and features. Write them by hand or have Claude build them.
Hermes and Claude Code are both agent runtimes; mods are the same extensibility layer we built with skills, but shipped as a first-class plugin system.
claude.dev launches
Anthropic launched claude.dev: developer docs, performance deep dives (made the web app 3x faster), Opus 5.5 effectiveness guide, eval hillclimbing.
Performance optimization and eval automation guides: both apply to Chrysalis and our own agent evals.
Fine-tuned Qwen3 4B beats Claude Sonnet 4.6
Fine-tuned Qwen3 4B on AWS outperforms both the base model and Claude Sonnet 4.6 on a specific task, cheaper and faster.
We run Sonnet for SalesKick grading; task-specific fine-tuning at 4B could drop cost and latency if the quality holds on our own evals.
Python releases: nothing in the window.