AGENTS·APR 24·11 MIN
Tokens as a budget: SESSION_STATE and the paste handoff
A dialogue on running multi-day Claude Code builds without trashing your context. The SESSION_STATE pattern, the paste handoff, and when to compact.
READ →
Topical cluster
AI Agent Engineering
Everything about building production AI agent workflows. When to reach for a sub-agent versus a skill versus an MCP server, how to think about context window economics, what parallel agent dispatch actually costs, and the failure modes nobody mentions in the marketing deck.
14 postsFor: Engineers and operators building AI agent systems for real work
Go deeper
AGENTS·APR 24·13 MIN
Before 60 writer agents ship: grep your MDX component props
An MDX component audit I now run before dispatching parallel writer agents. Five greps catch the prop drift and parse hazards that shipped silently.
READ →
AGENTS·APR 16·8 MIN
Building your first Claude skill: the five-file template
A hands-on tutorial for building your first Claude skill. Five files, one working skill, every decision explained without the theory.
READ →
AGENTS·MAR 23·7 MIN
Token budget math for Claude Code: what 104K per task means
Field notes on Claude Code token budgets. The 200K context is not 200K of workspace. Real numbers on system prompt, tool defs, and working set.
READ →
AGENTS·MAR 23·9 MIN
When to spawn a Claude Code sub-agent versus a tool call
A decision log for when to spawn a Claude Code sub-agent versus just making a tool call. Four signals, three real examples, and the dispatch math.
READ →
AGENTS·FEB 25·7 MIN
Prompt caching economics: when the 5-minute TTL pays rent
Field notes on Claude prompt caching economics. Real numbers on the 5-minute TTL, 1-hour option, and the looped workflows where caching is rounding-error.
READ →
AGENTS·FEB 13·8 MIN
Agent teams for code review: the parallel-reviewer pattern
A pattern library for running parallel agent code review. Three reviewers, one coordinator, and the file-ownership rule that makes it reliable.
READ →
AGENTS·FEB 13·7 MIN
Agent failure modes and the recovery patterns that keep shipping
Field notes on Claude Code agent failure modes. Four repeatable failure shapes, the signals that tip me off early, and the recovery pattern for each.
READ →
AGENTS·FEB 7·9 MIN
MCP server architecture: when to build one versus use a skill
A pattern library for MCP server architecture. Three transport choices, four real workload shapes, and the test for when MCP is the wrong tool.
READ →
AGENTS·JAN 29·8 MIN
Parallel versus sequential agent dispatch: the real tradeoffs
A decision log for parallel versus sequential agent dispatch. When parallel wins, when it loses, and the coordination cost nobody mentions.
READ →
AGENTS·JAN 29·7 MIN
Background agents with run_in_background: when it pays off
A tutorial on Claude Code background agents via run_in_background. When to use it, what breaks in foreground, and the three commands I always background.
READ →
AGENTS·JAN 12·8 MIN
MCP vs skills vs tools: picking the right Claude extension
A decision log for MCP vs Claude skills vs plain tool calls. Six criteria, three columns, and the lifecycle test that settles most picks.
READ →
AGENTS·JAN 9·9 MIN
Git worktrees for parallel agents: the isolation pattern
A pattern library for using git worktrees to isolate parallel Claude Code agents. Three instances, the setup script, and the file-ownership rule.
READ →
Why this matters.
AI agent engineering in 2026 is past the demo phase. Claude Code sub-agents, MCP servers, and the skill system are stable enough to ship into production, and most of the interesting design questions are not 'can the agent do this' but 'should this be a sub-agent or a skill, and what is the context window economics of the choice'. A sub-agent has its own context budget and a clean exit; a skill loads inline and shares the parent's tokens; an MCP server runs out of process and persists across sessions. Picking the wrong shape can triple your token spend without changing the user-facing behavior.
This cluster documents the patterns I use when building agent workflows for real work. Parallel sub-agent dispatch with file-ownership boundaries so two agents do not stomp on the same code. Skills versus tools versus MCPs, with the decision rubric. Context-window pressure tactics. Failure modes nobody covers in the launch posts: silent context truncation, sub-agent loops, skill discovery costs.
If you are building agent-driven systems for production, start with the orchestration hub piece and walk the sub-agent and skill posts before architecting the next workflow.
Put this to work
Claude Code sub-agents, MCP servers, skills, and the orchestration stack.
> See the Operator's StackOther clusters
Attribution
Attribution & CAPI
Server-side tracking, dedup, and the math behind DTC attribution.
Shopify
DTC Shopify Infrastructure
Theme architecture, metafields, agent-orchestrated builds, Hydrogen decisions.
Healthcare
Healthcare & Compliance
HIPAA-aware Next.js, audit trails, and regulated DTC patterns.
Solo brand
Creative-Tech Solo Brand
The hybrid creative-director-who-codes operating system.
Pricing
Services Business & Pricing
Productizing services, pricing strategy, and the retainer exit.
Lifecycle
Email & Lifecycle Marketing
Klaviyo flow architecture, retention math, and the LTV playbook.
Analytics
Analytics & Data Infrastructure
GA4, BigQuery, and the warehouse-first analytics rebuild.
Conversion
Ecommerce Conversion & UX
PDP patterns, cart decisions, checkout extensions, and the CRO stack.
Programmatic SEO
Programmatic SEO & Content Ops
Scaling content without burning author brand authority.
Paid Social
Paid Social Performance
Meta, TikTok, and Google Ads for DTC operators who also own the data.
Brand
Brand Architecture & Design Systems
Visual identity, naming ladders, and design systems that scale with the business.
Shopify Apps
Shopify App Ecosystem
Selecting, configuring, and replacing the third-party stack on Shopify.
Fractional
Fractional Ops & Service Leadership
Running a high-leverage fractional practice without becoming an agency.
Image AI
Local AI Image Generation
Running Flux, Z-Image, and Qwen locally without the cloud-API bill.
Let’s fix
some problems.
Three short steps below. I read all of these, it’s just me on the inbox. Usually you get a real reply within a day, sometimes the same day if I’m at the desk.
or email direct hello@michaeldishmon.com
01 / 03WHAT YOU NEED
$ cat lead.json | mail -s 'new signal' michael
