Voice agents in production: what to automate first
Automate predictable, recoverable calls first. Keep complaints, emergencies, and irreversible decisions with people.
ReadAll topics
The 3 most used TypeScript agent frameworks in 2026
npm downloads put the Vercel AI SDK ahead of every other TypeScript agent framework. The August 2026 ranking, with the caveats it deserves.
Read
The 6 most used Python AI agent frameworks in 2026
PyPI downloads put LangChain far ahead of every other Python AI agent framework. The August 2026 ranking, with the caveats it deserves.
Read
Don't hire until your operations are designed
We nearly hired a marketer and a salesperson. The system they would have joined did not exist yet, and a hire cannot fix that.
Read
What AI summaries drop, and why it cost us 3 times
Summaries are lossy in one direction: they drop reversals, negative findings and non-events, the class of fact that changes a decision.
Read
Context engineering in deep agents: the four levers
A deep agent's ceiling is its context window. The harness ships four separate levers, and knowing which one is working decides whether a long run finishes.
Read
Deep agents in production: what you still own
LangChain's deepagents ships subagents, a filesystem, and summarization by default. You still own the human gate, the evals, and the cost.
Read
How we halved Claude Code token usage with RTK AI
We routed our coding agents' shell commands through RTK AI, an open source CLI proxy. Verified: 49.6% of tokens saved across 1,456 commands.
Read
Stop gating on LLM confidence: what to use instead
Self-reported LLM confidence clusters near certainty. Compute a signal you can measure, then escalate the exceptions to a person.
Read
Coding is no longer the bottleneck. Trust is.
AI made writing code cheap. Architecture, verification and security did not get cheaper, and that is where the real work moved.
Read
Codex CLI: which commands you can actually build on
Three May releases pointed somewhere; 35 more confirmed it. The signal is not the changelog, it is the maturity label on each command.
Read
LangGraph human approval: interrupt or middleware?
LangChain ships tool-call approval as middleware now. When you still need interrupt() by hand, and the replay rule that bites either way.
Read
Two Claude Code fixes are off by default: pin them
One of them drops your unpushed commits from new worktrees. What changed under you, and the settings that hold it.
Read
Hermes Agent memory: a hard cap, not an archive
It refuses any write over its context budget and runs a curator that ages self-written skills out. What an agent keeps, and who prunes it.
Read
The 7 Claude Code plugins we keep installed
After testing 80+ official Claude Code plugins, these are the 7 that changed how we ship software every day.
Read
LLM-as-judge in LangSmith: calibration is the work
Wiring an LLM judge takes an afternoon. Calibrating it against human labels is what decides whether its scores mean anything at all.
Read
Token attribution in LangChain: tokens are not costs
Two callbacks in langchain-core attribute LLM usage per user, workflow or agent node. The streaming opt-in is now a default, and a token count is not a bill.
Read
Prompt versioning in 2026: the commit model won
Providers now version prompts as immutable commits. One is retiring its hosted registry, so depend on the commit model, not on the registry.
Read
Prompt caching across Claude, GPT and Gemini
Cache reads cost a tenth of fresh input tokens. The providers have converged on the same economics, and the old implicit-versus-explicit split no longer holds.
Read
LangChain model profiles for context management
A hardcoded token threshold is a model-specific assumption in disguise. LangChain's model.profile lets context logic read the real limits at runtime.
Read
LangGraph agent skills: decisions, not snippets
We rebuilt our six-day research agent against current LangGraph. Every architecture decision the skills made held up; four of their code samples no longer run.
Read
LangChain agent skills for AI coding assistants
Your assistant knows general programming, not your framework's production patterns. Agent skills close the gap, loading a pattern only when it is needed.
Read
Agent Lightning 1.0 has no APO: what we kept
We optimised a LangGraph agent's prompt with Agent Lightning. The algorithm has since left the framework; the lesson it taught has not.
Read
AI framework vs custom stack: the layer you own
The choice is not framework or custom. It is which layer you keep the right to replace, and that is what decides if a migration hurts.
ReadNo matches