The Pulse — September 15, 2026
Anthropic's Claude now writes 80% of their code, with engineers shipping 8x more per quarter. Meanwhile, the agent memory gap is getting defined: three new papers benchmark cross-session risks, inference budget constraints, and multi-agent orchestration trade-offs we're hitting in production.
Models
DeepSeek Pro Latest and Flash Latest on OpenRouter
DeepSeek's latest routing handles on OpenRouter: Pro at $0.66/$1.98 per million tokens, Flash (multimodal) at $0.15/$0.60, both with 1M context.
Competitive long-context pricing for retrieval and document agents where we route to the cheapest capable model.
Research
MemRiskBench: Trace-Aware Risk-Preserving Evaluation for Long-Horizon LLM Agents
Benchmark for memory-specific risks in long-running agents: stale facts, conflicting updates, cross-user leakage, and constraint decay, with per-risk failure rates instead of aggregate scores.
Addresses exactly the durability and isolation issues we're architecting around in Pipeline's multi-tenant voice agents.
GraMRAG: Orchestrating Multi-Agent Multi-Step Reasoning via Graph Memory with Reinforcement Learning
Graph-structured memory for multi-agent RAG with RL-driven orchestration; addresses state blindness in knowledge-intensive multi-step tasks.
Relevant pattern for coordinating document agents and maintaining shared retrieval state across turns.
LIMBO: Lifelong Inference-Time Memory and Budget Optimization for LLM Agents
Balances what to replay from memory against per-call token budget; lifelong agents can't afford unbounded context growth.
The inference budget problem for stateful agents we hit every time context windows fill up mid-session.
Salesforce Koa: An Enterprise Language Model for Agentic Tool Use
Salesforce fine-tuned Nemotron-3-Super-120B with GRPO on public and synthetic data for tool use and agentic workflows; no customer data used.
Tool-calling benchmark and training recipe relevant for SalesKick's agent stack.
MTAC-IFBench: Benchmarking Instruction-Following in Multi-Turn Agentic Coding
Tests whether coding agents follow process constraints (no library X, test coverage above Y) across turns, not just functional correctness.
Exactly what we gate in PR review: adherence to repo rules and conventions, not just passing tests.
Releases
Langfuse v4.36.0
AI gateway now relays authenticated OpenAI Realtime API requests; filter UI moved to inline suggestions; session timeline toggle; configurable event propagation block sizes.
Langfuse is our observability layer; Realtime API support matters for Pipeline's voice agents, and event tuning helps with high-throughput logging.
Hacker News
Pion, an agent designed to run any company autonomously
Andon Labs' autonomous company management agent pitch: cross-functional orchestration, continuous operation, and delegated decision-making.
Worth knowing where the agent hype is pointing; no direct application yet, but instructive for scoping what 'autonomous' doesn't mean in practice.
OpenAI bots knew about the RubyGems caching vulnerability
Public post describing how OpenAI's web crawler accessed and likely indexed a disclosed but not yet widely announced RubyGems cache poisoning bug.
Reminder that scraping and training cut both ways: agents inherit knowledge, including security gotchas we haven't patched yet.
How to write an effective software design document
Practical guide: problem, constraints, options, decision, and why; skip the filler, write it before the PR, update it when reality diverges.
Reinforces the design-first workflow we're enforcing on larger features; agents can draft these now, and this is the template to hand them.
Notes on gotchas while migrating 35kb preprompts from Opus to self-hosted Ollama
Field notes from a real migration: chunking, quality drop on reasoning, latency trade-offs, and which open models held up under production load.
Relevant if we consider self-hosted fallback for high-volume low-margin endpoints; the gotchas are the actually useful part.
GPT-5.6 Luna vs. GPT-6 Astra: Is a $1.20 Model Good Enough for Code Review?
Benchmark comparing Luna and Astra on code review tasks: Luna acceptable for routine checks, Astra better for architectural insight and subtle bugs.
Directly applicable: we run wehshi with multi-model panels; knowing where cheaper models suffice saves us real money.
People on X
Addy Osmani on Anthropic's agentic coding scale
Addy Osmani on Anthropic's internal CI scaling: Claude writes 80% of the code, engineers ship 8x more per quarter, tests grew 10x, CI jobs up 25x in six months.
Real production numbers on agentic coding's blast radius; we're seeing the same test explosion and build time pressure.
shadcn/lint: agent-first Tailwind linter
shadcn/lint announcement: agent-first linter for Tailwind; errors explain what broke and how to fix it using your design system's components and theme.
Solves the agent-breaks-design-system problem we hit when agents freestyle CSS; constraint enforcement at lint time, not in PR review.
Cline desktop app launch
Santiago on Cline's desktop app launch: agentic coding tools moving away from TUIs toward desktop apps with better affordances for multi-file, multi-step workflows.
Pattern shift worth watching; we lean on terminal agents for CI and deploy, but richer UX might reduce the handoff friction.
Hermes Business accounts
Nous Research announces Hermes Business: team accounts with shared balance, per-member caps, and shared skills that compound into proprietary IP.
We're Hermes users; team accounts with shared skill libraries would centralize the runbooks and agent workflows we're building per client.
Models: quiet day, only DeepSeek routing handles worth carrying. Next.js and Python: nothing notable in the window. DSPy, LangChain, Node: silent. Google's DevFest announcement and OpenAI's Fyxer and Perplexity case studies were vendor posts without new technical surface.