Insights
Notes from production Claude work.
What we see inside repositories: agents that stall, evals that do not exist, tools that were never designed, and the fixes that get them onto real traffic.
The library is grouped the same way we staff work. Production systems, quality and cost, and shipping inside real companies. New notes land weekly.
Production systems
How Claude agents, tools, retrieval and architecture hold up once they leave a demo.
Claude Agent SDK Session Recovery: Resume, Fork, Compact, and Checkpoint Safely
Learn when to continue, resume, fork, compact, and checkpoint Claude Agent SDK sessions so long-running agents can recover safely.
Read the noteWhy most Claude pilots never reach production
Most Claude pilots stall on integration, evaluation and cost, not prompting. Here is what actually blocks a production launch, and how to unblock it.
Read the noteMCP versus ad-hoc Claude tools: pick the contract, not the acronym
MCP is a protocol. The production question is whether Claude has a narrow, permissioned contract with your systems of record.
Read the noteClaude agent architecture that survives real traffic
A production Claude agent is a loop with budgets, retries and checkpoints, not a chat window with tools attached.
Read the noteRetrieval for Claude agents: hybrid search, not a bigger window
Long context does not replace retrieval. Production agents need a small, cited, testable set of chunks and a budget for what enters the prompt.
Read the noteClaude tool schema design: small actions, loud errors
Tool schemas are the product Claude actually uses. Narrow the action, require the fields, and make failure obvious in traces.
Read the noteContext engineering for Claude: budget the window like memory
Context engineering is choosing what Claude sees, in what order, and what is cached. It is the discipline behind quality and spend.
Read the noteMulti-agent Claude systems: when a second agent earns its keep
Split agents when you have a real seam (research versus mutate, or two systems of record). Do not split them to look advanced.
Read the noteClaude computer use in production: isolate the desktop, log the clicks
Computer use is an interface of last resort. Sandbox it, constrain the goal, and prefer APIs or MCP when the system of record has one.
Read the noteQuality and cost
Evals, spend, latency and the gates that stop a quiet regression from hitting users.
Claude Agent Observability: Trace Model Calls, Tools, Cost, and Failures
A production guide to tracing Claude model calls, tools, cost, failures, and task outcomes with OpenTelemetry-shaped telemetry.
Read the notePut eval gates on your Claude agent before you scale traffic
An eval harness turns Claude quality into a CI number. Here is the minimum setup that catches regressions before customers do.
Read the notePrompt caching is how most Claude bills actually come down
Uncached long context is the usual 3x overspend. Cache the stable prefix, budget the rest, and measure cost per successful task.
Read the noteStructured outputs make Claude agents callable by software
If downstream code cannot parse the answer, you do not have an integration. Use schemas, validate, and eval the JSON, not the vibe.
Read the noteWhat a two-week Claude Expert Audit actually inspects
An audit is traces, spend, tools, safety and a ranked remediation plan you can retest. It is not a slide deck of prompting tips.
Read the notePut Claude eval gates in CI or accept shipping on instinct
If the golden set is not a pull request check, it will not run the week you are late. Block on safety and writes; warn on softer slices.
Read the noteClaude latency in production: timeouts, streaming and budgets
Latency is a budget. Cap steps, cache the prefix, stream when a human is waiting, and fail before the user does.
Read the noteShipping in companies
Safety, Claude Code, regulated deployments and the human loops that get a system live.
Prompt injection defense for Claude agents that can write
Injection is a product and permissions problem. Treat untrusted text as data, constrain tools, and put a human on irreversible actions.
Read the noteClaude Code workflows that your engineers will actually keep
Claude Code is useful when the repository has conventions, evals and a review path. Without those it is a faster way to merge guesswork.
Read the noteShipping Claude agents under HIPAA, SOC 2 and SOX
Regulated Claude work is a deployment and audit-log problem first. Model quality is necessary and still not sufficient for a security review.
Read the noteWhere to put a human in a Claude agent loop
Human review belongs on irreversible writes and ambiguous intent, not on every token. Design the queue like a product, not a guilt valve.
Read the noteClaude on Bedrock, Vertex or the Anthropic API: pick for control, not fashion
The right Claude runtime is about data residency, IAM and procurement, not a leaderboard. Match the platform to the review you have to pass.
Read the noteTrain your engineers on Claude using your codebase, not a sample app
Enablement that sticks is prompting, tools, evals and Claude Code on the repository you ship. Exam prep is optional. Habits are not.
Read the note