Monthly field guide

State

A working map of the AI coding stack: tools, models, frameworks, context systems, and the tradeoffs that actually affect shipping.

88 May signals4 core shiftsUpdated monthly

May 2026 Snapshot

Start with the current map

Read May issue

May issueState of AI Coding: What Changed This MonthMay's field note on control planes, CopilotKit UI contracts, Mastra workflows, security, and cost discipline.Stack mapThe New AI Coding Stack I Would Pick TodayThe layered stack map: terminal agent, editor, background agent, Mastra, CopilotKit, MCP, context, security, and cost.Change filterThe Model, IDE, CLI, and Agent Framework Changes That Actually MatterA decision filter for what to act on across models, editors, terminal agents, background agents, frameworks, context, security, and cost.Framework mapAI Agent Frameworks ComparedLangGraph, CrewAI, Mastra, CopilotKit, Vercel AI SDK, and where each one actually fits.Latest signalTaste Skills Are Turning Agent Review Into InfrastructureWhy review quality, design taste, and repeatable skills are becoming part of the coding-agent stack.Context layerLocal Code Graphs Are the Next Agent Context LayerThe next jump is not a bigger prompt. It is a better map of the repository the agent is editing.

Signal Map

Four shifts worth tracking

Frameworks are becoming product choices

The decision is no longer LangGraph versus everything else. The real split is orchestration, UI copilots, durable TypeScript agents, and simple SDK loops.

Agent Frameworks ComparedA practical decision map for LangGraph, CrewAI, Mastra, CopilotKit, OpenAI Agents SDK, and Vercel AI SDK.Start here Mastra vs CopilotKit vs LangGraphThe same agent-native product brief mapped across backend workflows, UI copilots, and graph control.Field note When CopilotKit Is the UI Layer, Not the Agent FrameworkWhy CopilotKit belongs at the product surface around Mastra, LangGraph, or another backend agent.UI layer Mastra for Durable TypeScript AgentsWhere Mastra fits when a TypeScript product needs workflows, memory, tools, MCP, evals, and traces.TS backend Build an AI Agent Web App with LangGraph and CopilotKitWhat changes when the agent has to live inside a real application interface.App pattern

Context is moving from prompt text to infrastructure

Coding agents need retrieval, repo maps, memory, and selective context. Dumping more tokens into the window is the weakest version of the architecture.

Coding Agents Need Codebase Maps, Not Bigger PromptsWhy codebase graphs are becoming a necessary agent primitive.Repo maps Claude Knowledge Work Plugins Turn Agent Setup Into Team InfrastructureThe setup layer is becoming something teams manage, share, and review.Team setup Agent Memory Benchmarks Are Not EnoughMemory needs provenance, deletion, and workflow fit. Benchmark numbers are not the whole story.Memory

Cost and review are now product constraints

The agent that writes code is only part of the system. The expensive parts are review, retries, context burn, and failed delegation.

AI Agent PMF Is a Cost Control Problem NowAgent products win when they turn unpredictable work into bounded spend.FinOps AI Code Review Is the New BottleneckGeneration got faster. The review queue became the constraint.Review Claude Code Token Burn Is an Observability ProblemAgent cost needs instrumentation, not vibes.Observability

Security is becoming an agent workflow problem

The important question is not whether agents can find security issues. It is whether teams can triage, constrain, audit, and roll back what agents do.

The Agent Security Checklist I Use Before Connecting ToolsThe preflight for tool access, MCP servers, files, approvals, logs, and rollback.Checklist Prompt Injection in Agent Apps: The Practical VersionWhy hostile web pages, issues, PDFs, and tool outputs become dangerous only when they can drive side effects.Threat model Permissions, Logs, and Rollback for AI Coding AgentsThe operating loop for making coding-agent work reviewable, reversible, and boring.Operating loop AI Security Scanners Move the Bottleneck to TriageMore findings do not help if the review system cannot sort signal from noise.Triage Approval Fatigue Is an Agent Security BugPermission prompts fail when they train humans to click through.Permissions Claude Code Plugin URLs Turn Skills Into a Supply ChainAgent skills and plugins need the same supply-chain discipline as code.Supply chain

Monthly Issue Format

The repeatable editorial loop

What changed

The concrete product, model, framework, pricing, or workflow change that matters this month.

Why it matters

The practical consequence for working developers. What gets easier, what gets more expensive, and what becomes safer.

What to try

One tool, setup, repo pattern, or comparison path to test before the next issue.

What to ignore

The noisy launch, vague benchmark, or product claim that is not worth rebuilding your stack around.

Next Issues

What this hub expands into

Framework field notes: the same agent app built with Mastra, CopilotKit, and LangGraph.
Pricing and control: how model routing, cache discipline, and review queues change the real bill.
Agent security watch: permissions, tool access, supply-chain risk, and triage workflows.

Get the next issue See current focus

Get Smarter About AI Dev

New tutorials, open-source projects, and deep dives on coding agents - delivered weekly.

One email per weekReal code, not theoryFree forever