Monthly Marks - November 2025
Links I found useful in November 2025.
API and Platform Design
Designing APIs reinforces what I keep learning the hard way: boring is good, avoid breaking changes, treat versioning as a last resort.
Workflow Runtimes
I’ve been looking at Temporal alternatives:
- Motia uses a “Steps” concept where one file yields API/queue/worker with tracing. Works across TS/Py/JS.
- Restate is a single-binary durable runtime focused on agents. Claims p99 <100ms and 13k workflows/s.
- DBOS backs everything with Postgres—durable workflows, exactly-once queues, cron, replayable traces.
Infra and Security
Ten Docker features you’re probably not using covers multi-stage builds, BuildKit cache/secret mounts, profiles, buildx, healthchecks, Scout SBOM, hardening flags. I wasn’t using profiles.
Dependency cooldowns makes the case for a 7-14 day delay on new package versions. Blocks most supply-chain attacks. Dependabot and Renovate both support this.
Data and Databases
Scaling pgvector - pre vs post-filtering can flip 50ms to 5s. The quantization notes (halfvec, bit indexes) are worth reading if you’re hitting performance walls.
Aurora RDS failover race describes a case where a reader briefly became writer. They detected it via read-only UPDATE errors and replica commit spikes.
RAG Techniques
- Advanced RAG Techniques - notebooks covering agentic and basic approaches.
- GraphRAG - Microsoft’s graph-based RAG pipeline with CLI. Watch the indexing cost.
Agent Tooling and Evaluation
- ADK for Go - code-first agents from Google. Gemini-optimized but model-agnostic. Cloud Run friendly.
- Braintrust - prompt/eval loop with scorer gates and prod observability.
- Strix - autonomous pen-test agents with PoC validation. Fails CI on findings.
LLM Releases
OpenAI:
As I was checking model cards going through this list I noticed OpenAI added a model compare tool: https://platform.openai.com/docs/models/compare - thats great for trying to understand differences between the growing list of models!! I just wish links would include the models you were comparing for easier sharing/documenting!
- GPT-5.1 family (gpt-5.1, -chat-latest, -codex, -codex-mini) with new reasoning stack
- GPT-5-Codex-Mini - compact/cost-focused variant
- GPT-5.1-Codex-Max - now default in Codex CLI
Anthropic:
- Claude Opus 4.5 - 200k ctx/64k out, tuned for coding and agents
Google:
- Gemini 3 Pro - multimodal
- Gemini 3 Pro Image (Nano Banana Pro) - image generation
Open models:
- Olmo 3 (32B “Think”) - fully open from Ai2 with interpretable traces
- DeepSeek-Math-V2 - IMO gold performance
- Kimi K2 Thinking - 1T MoE (32B active) from Moonshot