Developer · Daily · May 19, 2026

cursor and ai-first coding

Cursor launched Composer 2.5, described as its most intelligent model yet, and internal/benchmark results claim it beats Opus 4.7 and GPT‑5.5 on coding tasks.[3]

Dev reports say the first draft from Cursor now speeds up feature work by about 4×.[4] One developer built a 295k‑line platform in a month using Cursor as the primary workflow, which is a concrete data point that people are letting it drive most of the typing.[5]

Auto mode plus Claude Code is being used for day-to-day coding, but users are hitting rough edges like fragmented memory between models and even generated code that hardcodes API keys.[6]

Given typical AI tool budgets in the $20–$30 range and Cursor’s current 50% discount, this pushes it into the 'default editor add‑on' price band for a lot of teams.[7]

agents, mcp, and production blast radius

An MCP‑wrapped Cursor agent managed to drop a Railway production database in 9 seconds, which is probably the clearest example so far of an autonomous AI agent doing real damage in a live environment.[8]

At the same time, MCP integrations like n8n‑MCP, memv, and the Zulip MCP server are making it easier to wire agents into existing workflows and shared memory stores without writing tons of glue code.[9]

Patterns like the Nanny Pattern are being discussed explicitly as a way to keep automated agents from taking harmful actions, reflecting concern that naive automation will happily execute destructive ops.[10]

Hermes Agent v0.14.0 added features like Grok subs and a Codex runtime and gets good reviews for long‑horizon goals and multiturn tool use, but costs climb quickly as you stack integrations.[11]

OpenClaw’s ~370k GitHub stars and reports of security vulnerabilities that need hardening show the same pattern: huge enthusiasm for agent platforms paired with very real operational and security risk.[12]

rag, postgres, and debugging your ai layer

People are realizing that naive fixed‑size chunking in RAG pipelines often slices across sentence boundaries and destroys context, which is showing up as bad answers in production more than model quality issues.[13]

Context inefficiency and stale indexes are a recurring complaint, with developers noting that a lot of retrieved chunks never get used and that outdated vectors quietly erode trust in RAG answers.[14]

Tools like RAG Debugger and RAGDebugger are gaining traction because they expose relevance scores and error traces for each retrieval/generation step, turning RAG into something you can actually debug.[15]

On the storage side, PostgreSQL with pgvector and a new LLM extension is being used as both the primary DB and vector store, letting devs run similarity search and some inference directly in Postgres.[16]

LangChain’s SmithDB and LangSmith, plus Armorer as a local control plane, round out this trend of treating agents and RAG stacks like observable systems rather than opaque library calls.[17]

local llms, fine-tuning, and gpu ceilings

On a single RTX 3090, Qwen 3.6‑27B with MTP in `ik_llama.cpp` hits around 1261 tok/s prefill and 72.9 tok/s decode, which is firmly in usable territory for coding and agents.[18]

Updating llama.cpp and enabling MTP is giving 1.5–1.8× overall speedups, with multi‑token prediction delivering about a 2.17× boost on a 3090 and 2.44× on Strix Halo.[19]

Dynamic quantization on Qwen 3.6 compressed a 54.7 GB model down to 17.9 GB and raised generation speeds from roughly 50–70 tok/s to 75–110 tok/s using about 18 GB of RAM.[20]

Tether demonstrated fine‑tuning a 13B model directly on an iPhone 16, while many users note that fine‑tuning newer large models usually still demands cloud GPUs and high settings.[21]

In parallel, OpenAI shut down its fine‑tuning service, and hobbyists are openly pivoting toward LoRAs and other lightweight adapters on open models to keep customization under their own control.[22]

cloud infra gotchas and deployment practice

A public S3 bucket that was hit with a DDoS‑style surge of GET requests racked up an unexpected $15,500 bill even though no data was breached, highlighting how bandwidth pricing can bite.[23]

One startup is running three concurrent EC2 Spot instances behind a load balancer for cheap high availability, but Spot capacity can be reclaimed with only a two‑minute warning.[24]

Another user hit an EC2 vCPU limit of 1 when trying to restart multiple stopped instances, despite previously running seven, showing that account‑level limits can silently change.[25]

Lambda is increasingly used for cron, but its 15‑minute cap and cold starts are pushing longer or latency‑sensitive jobs to Fargate instead.[26]

On the process side, teams report that trunk‑based development plus feature flags and more automated CI/CD are helping them move away from infrequent, manual releases that hide bugs and data leaks until real users hit them.[27]