Production Practice

Capacity Planning for LLM Apps: Rate Limits, Concurrency, and the Queue You Already Have

October 12, 2026

Your app does not fall over at high load because the model is slow. It falls over because nothing in the stack bounds concurrency, and provider 429s arrive as user-visible errors. Here is how to size, queue, and degrade on purpose.

Parsing PDFs for RAG: Layout, Tables, and the Pages Your Pipeline Silently Drops

October 5, 2026

Retrieval quality has a floor set by your parser. How to audit extraction, route documents by class, keep tables intact, and gate parse quality at ingest.

Subagents Without the Sprawl: When to Split a LangGraph Agent, and What It Costs

September 21, 2026 · TokenTailors

Multi-agent architectures are the default answer right now. Most teams need one or two subagents, not a org chart. Here is how to decide, and how to keep the split from eating your latency and your debuggability.

Latency Budgets for LLM Apps: Where the Seconds Actually Go

September 21, 2026 · TokenTailors

Time to first token, stage-level timing, prefix caching, parallel retrieval, and streaming: how to find and cut the seconds in a production LLM app without trading away quality.

Permission-Aware RAG: Enforcing Document Access Control in Retrieval

September 18, 2026

The moment your corpus spans more than one customer or one security group, retrieval becomes an authorization surface. How to filter at the index, keep recall honest, and test for leaks.

Migrating to LangChain 1.0: create_agent, Middleware, and the Code You Get to Delete

September 17, 2026

A staged migration off AgentExecutor and legacy chains: what create_agent and agent middleware actually change, what to delete instead of port, and how to guard the refactor with evals.

Swapping Embedding Models Without a Retrieval Regression

September 16, 2026 · TokenTailors

A new embedding model tops the leaderboard and someone proposes a re-index. Here is how to decide whether it helps your corpus, and how to run the migration behind a shadow index without a quality cliff.

Structured Output That Holds in Production: Schema Design, Validation, and Repair

September 15, 2026 · TokenTailors

Constrained decoding stops malformed JSON. It does not stop wrong JSON. How to design extraction schemas, validate beyond the type system, and repair failures without a retry loop that hides them.

Indirect Prompt Injection Is a Data Problem: Containing Untrusted Text in LangGraph Agents

September 14, 2026

Every retrieved document and tool response is untrusted input. Provenance tracking, capability limits, deterministic action gates, and the injection evals that tell you the defenses hold.

Context Engineering for Long-Running Agents: What to Keep, Compact, and Throw Away

September 11, 2026 · TokenTailors

An agent that works for five turns and degrades by thirty has a context problem, not a model problem. Budgets, tool-output pruning, compaction, and external memory in LangGraph — with the evals that tell you it worked.

Wiring MCP Tools Into a LangGraph Agent Without Losing Control

September 10, 2026 · TokenTailors

MCP makes it trivial to hand an agent thirty tools. Tool selection accuracy, schema discipline, permission boundaries, and how to measure whether the agent calls the right thing.

Cutting LLM Spend Without Cutting Quality

August 11, 2026

Most LLM cost problems are visibility problems. Per-feature token accounting, caching, model routing, and context discipline — with evals proving quality held.

When to Take LangChain Out of the Stack

July 20, 2026

A framework decision is an engineering decision, not a loyalty test. The signals that LangChain is costing more than it saves — and how to exit safely where it is.