Everyone can define the word. Almost nobody can point at the part that broke.
An agent has six parts, all of them published by the labs that build them, and once
you can name the six you can diagnose a bad output in about ten seconds instead of retyping
your request in a louder voice. This lesson is the anatomy, the loop it runs, the spectrum
it sits on, and the fault map that turns all of it into a working habit.
Updated August 1, 2026Refreshed monthlySources: Anthropic · OpenAI · Google
Core concept 01
Six parts, and you can point at every one.
Anthropic defines agents as models that direct their own process and use tools to get a
job done. OpenAI's guide for businesses breaks one down into a model, instructions, and
tools. Google's agents whitepaper adds the piece that makes it move: an orchestration layer,
the loop. Put the three descriptions side by side and the same six parts show up every time.
Instructions are the standing rules you write once. Context is everything it
can see at this moment. The model is the part that reasons. Tools are what it
can actually touch, like search, your files, a draft in your inbox. The loop is the
orchestration that keeps it going. You are the checkpoint, and the labs treat that as
a component, not a courtesy: OpenAI's guide gives human intervention its own section in the
guardrails chapter.
Naming the six is the entire point of this lesson. When output comes back generic, or
confident and wrong, or beautifully formatted and aimed at the wrong task, one specific part
failed. People who can name the part fix it in one move. People who cannot retype the request
in a louder voice and hope.
The six parts · every one of them is somewhere you can actually interveneCore concept 02
Two parts decide what it knows. Two decide what it can do.
Instructions and context are the knowing half, and they are not the same thing.
Instructions are standing: the role, the rules, the way you sound, written once and
read before every message. Context is right-now: the files you attached, the thread so
far, what a tool just returned. Instructions are the job description. Context is the desk.
Almost every "the AI is bad at this" moment is one of these two, and telling them apart is a
real diagnostic. Output that is fluent and generic means thin context, because it had nothing
of yours to work from. Output that ignores a rule you definitely wrote means the instruction
was buried, contradicted, or three pages back. Neither one is fixed by asking more politely.
Go deeper: why long conversations get forgetful
Context is a fixed budget, measured in tokens, and everything competes for it: your
standing instructions, every file you attached, the entire conversation so far, and the
output of every tool call the model made. When the budget fills, something has to give.
Products handle this by summarizing or dropping the oldest turns, which is why a long
chat starts losing details you gave it at the top.
The operator move that follows: a fresh conversation with the right five files beats a
forty-message thread almost every time. Start over more often than feels natural. You are
not losing progress, you are clearing the desk.
The model and tools are the doing half. The model is the reasoning: which
version you are talking to changes how well it plans a multi-step job, and most people never
look at the model picker in their own settings. Tools are the reach. A model with no tools can
only produce text about your work. A model with search, your files, and a drafts folder can
do the work and leave it where you will find it.
Go deeper: what MCP actually standardizes
Before the Model Context Protocol, connecting a tool to a model meant a bespoke
integration per tool, per model. MCP is an open standard, started by Anthropic and now
adopted across the industry, that defines one way for a model to discover and call an
outside tool. Build the connector once and any MCP-speaking model can use it.
Why a non-engineer should care: it is the reason the "connect your apps" menu in these
products keeps getting longer without anyone shipping a new model. Tool reach and model
smarts now improve on separate clocks.
Give a model a goal instead of a question and the loop starts: plan one step, act, look at
what came back, decide whether the job is done. Same four moves, over and over, with what it
just learned folded into the next pass. That is Google's orchestration layer, and it is the
only structural difference between a chatbot and an agent.
Concretely: "find three podcasts my ideal client listens to, draft a pitch for each, and
leave the drafts in my inbox" is a goal. The loop is the searching, the reading, the drafting,
and the rereading of its own weak draft before you ever see it. The four stations below are
worth memorizing, because every one of them can fail differently, and the fix depends
on which one did.
Go deeper: why the loop is the expensive part
Each pass through the loop re-reads the accumulated context, so a job that takes twelve
turns is not twelve times one turn. It is closer to compounding: turn twelve carries
everything from turns one through eleven. This is why an agent that wanders costs
noticeably more than one with a tight goal, and why "when are you done" is a real design
question rather than a philosophical one.
Practical version: scope the goal so the finish line is checkable. "Draft three pitches
and stop" ends. "Improve my podcast outreach" does not.
The loop in motion · plan, act, observe, decide, and the return that makes it an agentCore concept 04
Most of what gets sold as an agent should be a workflow.
Autonomy is a dial, not a switch, and there are three settings on it. In chat you pick
every step. In a workflow you picked the steps once, in advance, and they run on a
trigger from then on. In an agent the model picks the steps as it goes. Anthropic draws
this exact line for its own developers and then gives advice that most marketing skips: find
the simplest thing that works, and add autonomy only when it earns its keep.
That advice is worth actual money to you. The wins sitting in your business this week are
the inquiry replies, the follow-ups, the listings, the invoices, and those are workflow-shaped
almost without exception. They happen the same way every time, which is exactly the condition
under which paying for a model to rethink them from scratch is a waste. Knowing when the
powerful tool is the wrong tool is what engineers are actually paid for, and it transfers
directly to running a business.
Go deeper: the augmented model, and why the labs start there
Anthropic's building block for all of this is what it calls the augmented LLM: a model
plus retrieval, tools, and memory. Everything above that, including full agents, is
composition on top of the same block. The reason this matters to a non-engineer is that
it explains the recommendation: since autonomy is added on top, it can also be left off,
and leaving it off costs you nothing except a capability you were not going to use.
The failure this prevents is specific and common. Someone hands an autonomous agent a
job that runs identically every Tuesday, then spends a month debugging variance they
introduced on purpose.
The autonomy dial · chat, workflow, agent, and where the wins actually sitCore concept 05
Bad output is a diagnosis, not a verdict.
This is where the anatomy stops being trivia. Every characteristic failure maps to exactly
one of the six parts, and the fix is different for each. Read the left column the next time
something comes back wrong, and fix the part instead of the prompt.
The fault map · symptom, part, fix
It did this
The part that failed
The move that fixes it
Fluent, competent, and generic. Could be about anyone's business.
02 · ContextIt had nothing of yours to work from.
Feed it three real artifacts: your best past example, your actual numbers, your own words. Specificity in beats specificity out.
Confidently stated something that is not true.
02 · ContextYou asked it to recall instead of read.
Give it the source rather than asking from memory. Anything checkable should arrive with the thing to check against.
Ignored a rule you know you wrote.
01 · InstructionsThe rule was buried or contradicted.
Shorter standing rules, hard constraints first. Two rules it follows beat nine it averages.
Did the wrong task, beautifully.
01 · InstructionsNobody said what done looks like.
Name the finished artifact before it starts: "a 120-word email, one link, ready to send."
Said it would do the thing and then did not do it.
04 · ToolsIt had no reach, only words.
Check what it is actually connected to. A model with no tools can only ever produce text about the work.
Stopped halfway, or went in circles.
05 · The loopNo finish line, or the desk filled up.
Split the job into two smaller runs with a checkable end. Start the second in a fresh conversation.
Did something you would never have approved.
06 · YouThere was no checkpoint.
Put the approval gate before anything that sends, spends, publishes, or deletes. Draft, do not send, stays the house rule.
Two of the seven rows are context problems and two are instruction
problems, which is the honest shape of this. The overwhelming majority of disappointing AI
output is a briefing failure, not a capability ceiling, and briefing is a skill you can be
good at by Thursday. That is Lesson 02.
Fresh from the lab
What changed this month
OpenAI built a model that spawns its own subagents mid-task. On July 9, GPT-5.6 Sol
reached general availability with a new mode called Ultra, built into the model itself. Instead
of working a hard problem in one long reasoning chain, Ultra breaks the job into pieces, runs
subagents on several pieces at once, then merges what comes back, the same plan-act-observe loop
from this lesson's anatomy, just fanned out instead of run once. On Terminal-Bench 2.1, a test
that drops models into real terminal environments and scores whether they can finish
command-line jobs, Sol alone scores 88.8%. Sol with Ultra mode scores 91.9%. There's no separate
sticker price for Ultra, each subagent burns its own tokens, so the bill grows with how many
pieces get spun up.
The protocol connecting agents to their tools just got rebuilt underneath everyone. On
July 28, Anthropic shipped MCP 2026-07-28, the newest Model Context Protocol spec, and the
headline change is a stateless core: MCP moves from holding a live connection open to plain
request and response, the same shape as most web APIs. That lets MCP servers, the connectors
that give an agent its tool reach, run on serverless and edge infrastructure instead of needing
something always on. Anthropic says monthly SDK downloads for MCP just passed 400 million, four
times where they stood at the start of the year, and Claude's own connector directory now lists
more than 950 MCP servers.
A free-to-download model just closed the gap on tool use specifically. Moonshot AI
released Kimi K3 on July 16, a 2.8 trillion parameter model with a context window over one
million tokens, and made the weights free to download on July 26. On real-world task automation
benchmarks, tests that check whether a model can actually finish a job using tools rather than
just describe one, K3 finished first in four of eight categories, including AutomationBench and
BrowseComp, and scored 88.3% on that same Terminal-Bench 2.1, a hair behind Sol's 88.8%. The gap
between what you pay for and what you can download for free is shrinking fastest on exactly the
tool-use skills this lesson is about.
Each one is tied to a part of the
anatomy, so the vocabulary and the diagram reinforce each other. Learn these ten and every AI
conversation, pricing page, and settings menu gets easier to read. There is no quiz.
LLM
Large language model. Part 03. The engine inside Claude, ChatGPT, and Gemini. It
learned patterns from mountains of text and uses them to write, reason, and plan.
Prompt
Parts 01 and 02. Everything you hand the model in one message: the ask, the
background, the examples. Lesson 02 is entirely about writing better ones.
System prompt
Part 01. The standing instructions read before every message, rather than typed
into each one. In Claude Projects and custom GPTs this is the "instructions" box.
Context window
Part 02. The fixed budget of everything the model can see at once: instructions,
files, the thread, tool results. When a long chat gets forgetful, the window is why.
Token
How models count text, roughly three-quarters of a word each. Context windows, limits,
and prices are all measured in tokens, which is why the word is on every pricing page.
Tool use
Part 04. The model reaching beyond chat: searching the web, reading your files,
running code, drafting in your inbox. Tools turn a talker into a doer.
MCP
Model Context Protocol. Part 04. An open standard, started by Anthropic and now
used across the industry, that lets one connector plug a tool into many models. Think
USB, but for AI.
Agent
Part 05 is what makes it one. A model using tools in a loop toward a goal,
checking its own work as it goes. The subject of this lesson.
Human in the loop
Part 06. A review step placed in the design on purpose, not whenever you
remember. OpenAI's agent guide treats it as a guardrail, which is the right altitude.
Hallucination
A confident wrong answer. Models fill gaps with plausible text, which is why the fault
map sends this one to context: give it the source instead of asking it to recall.
Same anatomy, your business. Every
workflow below builds a standing assistant, a Claude Project or a custom GPT, aimed at the
task eating the most of your week. Budget about an hour, once. A free Claude or ChatGPT
account is enough to start.
Notice what the steps are actually doing, because
it is the same six parts in the same order. Writing the job description is part 01.
Uploading your real materials is part 02. Choosing Claude or ChatGPT is part
03. Deciding whether it gets to search or touch your files is part 04. Running
this week's real work through it and correcting the misses is part 05. Keeping the
final say is part 06, and it stays yours permanently.
For female founders
I'm the CEO, the marketer, the bookkeeper, and the intern, and the
building only happens after 9pm.
The everything-hat assistant
Tool: Claude Projects or a custom GPT · free tier works
Pick its one job. Not "help with my business." Pick the biggest single drain,
usually writing: proposals, follow-ups, social posts, or all three.
Create a Project and write its job description in the instructions: who you are,
who your customers are, how you sound, what a great draft looks like.
Feed it your real materials. Your best proposal, your services page, three
emails you were proud of. It writes like you when it has read you.
Run this week's actual work through it. When a draft misses, say why in plain
words. Two corrections in, the drafts start sounding like yours.
Make it the default. New inquiry, new post, new proposal: it drafts first,
you edit second. Drafting from zero is no longer your job.
The payoff: first drafts stop costing you evenings. The hour you
spend setting this up comes back the first week.
For artists
I make the work. Then the statements, applications, and pitches eat the
studio time the work needed.
The studio writing assistant
Tool: Claude Projects or a custom GPT · free tier works
Gather your raw materials: your current artist statement, a CV, notes on your
practice, and honest sentences about what your work is about.
Create a Project with instructions like: "You help me write about my work. My
voice is plain and specific. Never use words like 'explore' or 'juxtapose' unless I do."
Upload the materials. Add images' descriptions in your own words; how you talk
about the work is data.
Point it at a real deadline: the grant application or open call sitting in your
tabs. Have it draft the statement and answers from your materials.
Keep every final version in the Project. Each application makes the next one
faster, because it learns from what you actually submitted.
The payoff: applications stop being a lost studio day and become an
edit session with coffee.
For actors
Between survival jobs, self-tapes, and submissions, the admin of the
career crowds out the acting.
The submissions sidekick
Tool: Claude Projects or a custom GPT · free tier works
Load your kit: current resume, bio in three lengths, rep info if you have it,
and the types you're actually right for, in your own blunt words.
Create a Project instructed to write cover notes, slate scripts, and pitch
emails in your voice, warm and specific, never desperate.
For each submission, paste the breakdown and ask for a tailored cover note plus
which of your credits to lead with. Edit, send, move on.
Use it for sides prep: paste the scene and ask for given circumstances, beat
breakdown, and three playable choices. It's a scene partner that never gets tired.
Log what you booked and what you didn't back into the Project. Patterns show up
fast when something else is keeping the spreadsheet.
The payoff: submissions in minutes instead of an evening, and prep
that starts at choice three instead of a blank page.
For musicians
I'm a musician, and somehow most of my week is a promo job I never
applied for.
The release-day roadie
Tool: Claude Projects or a custom GPT · free tier works
Load the story once: your bio, past press, how this record was made, and the
five details fans always ask about.
Create a Project instructed to write like you talk from stage: pitch emails,
captions, newsletter blurbs, EPK copy.
Give it one release and ask for the full kit: playlist pitch, three venue
emails, a week of captions, and a newsletter, all from the same story.
Tune the voice out loud. "Too polished, I'd never say 'sonic journey'" is a
perfectly good correction, and it works.
Reuse the Project every release. The kit gets faster and more yours each time,
because your corrections stay in the room.
The payoff: release week promo drops from days to an afternoon, and
the writing time goes back to the songs.
For fitness pros
I'm booked teaching all day, so programming, check-ins, and content
happen in the cracks that used to be my rest.
The back-office training partner
Tool: Claude Projects or a custom GPT · free tier works
Load your method: how you program, your movement standards, a sample week for
two or three client types, and your check-in style.
Create a Project instructed to draft programs and check-in replies your way,
and to never invent exercises you don't coach.
Run Sunday programming through it: paste each client's week ("knee cranky,
travel Thursday") and have it draft the adjustments for your review.
Draft check-in replies from your quick notes. You approve every one; warmth is
yours, typing is not.
Batch content from what you already say: have it turn this week's cues and
client wins (no names) into a week of posts.
The payoff: Sunday admin shrinks to one focused hour, and clients
still get you, not a bot.
For coaches
I sell transformation, but my calendar is full of the sessions, so the
marketing that fills next month never happens.
The between-sessions content engine
Tool: Claude Projects or a custom GPT · free tier works
Load your frameworks: your niche, your method, client questions you hear weekly,
and two pieces of content that performed.
Create a Project instructed to turn themes into posts and emails in your voice,
with a hard rule: no client details, ever, even anonymized ones you're unsure about.
After sessions, jot one line about the theme ("everyone's negotiating in
August"). Friday, hand it the week's lines and get a content week back.
Have it draft the follow-ups too: discovery-call recaps and check-in emails
from your bullet points, ready to personalize.
Keep winners in the Project. What lands with your audience becomes the model
for the next batch.
The payoff: marketing happens every week, in the margins, instead of
never, in the someday.
For therapists
Between sessions I'm doing intake calls, insurance letters, and website
updates, and none of it is why I trained.
The practice-admin assistant (with a bright red line)
Tool: Claude Projects or a custom GPT · free tier works
Draw the line first: no client information goes into a general AI tool, period.
This assistant handles the practice, never the people. (Clinical notes need a
HIPAA-compliant tool with a signed BAA; that's a different aisle.)
Create a Project for the public-facing practice: your modalities, who you serve,
your fees and policies, your warm-but-clear tone.
Have it draft the repeatables: intake FAQ replies, waitlist emails, superbill
explainers, out-of-network benefit scripts clients can read to their insurer.
Refresh your website and directory profiles: have it rewrite your Psychology
Today profile from your instructions, then edit until it sounds like your office feels.
Build the referral kit: a clean one-pager about your practice for colleagues,
drafted, then made yours.
The payoff: the admin hour between clients goes back to notes, lunch,
or nothing at all, and the red line keeps it ethical.
For authors
I want to be writing the next book, and instead I'm writing about the
last one, forever.
The platform ghostwriter (for everything but the book)
Tool: Claude Projects or a custom GPT · free tier works
Decide what stays yours. The book is you. The synopsis variants, newsletter,
and event blurbs are the assistant's department.
Create a Project loaded with your book's synopsis, sample chapter, reviews,
bio, and the way you talk to readers.
Have it produce the eternal kit: synopsis in three lengths, five pitch angles
for podcasts, event copy, and a reader-magnet description.
Batch the newsletter: bring it your month's notes and let it draft; your edit
pass keeps the voice, the blank page stops being yours.
Point it at outreach: tailored podcast and book-club pitches from its stored
knowledge of your book, ten minutes each instead of ninety.
The payoff: platform upkeep drops to an hour a week, and the next
book gets the mornings back.
For chefs
I'm cooking or shopping all day. Menus, costing, and captions happen
after midnight if they happen at all.
The back-of-house brain
Tool: Claude Projects or a custom GPT · free tier works
Load your kitchen: signature dishes, sourcing philosophy, price points, menu
voice ("no adjectives I wouldn't say out loud").
Create a Project that writes menu descriptions, catering quotes, and specials
posts the way you'd say them across the pass.
Run a real menu change through it: paste the new dish and get the menu line,
the caption, and the newsletter blurb in one go.
Draft catering quotes from your numbers: give it your per-head costs and
margins once, and it assembles the quote email while you prep.
Keep a specials bank. Every special it writes up goes back in, so January's
menu writes faster than July's did.
The payoff: the after-midnight admin shift gets cut from the
schedule, and the menu still sounds like you.
For consultants
I'm delivering client work all day, so my own pipeline goes cold every
time I'm busy, which is exactly when it shouldn't.
The pipeline keeper
Tool: Claude Projects or a custom GPT · free tier works
Load your positioning: who you serve, the three problems you solve, a winning
proposal, and two case studies (client names optional).
Create a Project instructed to draft proposals, follow-ups, and thought-leadership
posts in your voice, credible and unhurried.
After every call, give it your notes and get the recap email and next-step
draft back before the coffee's cold.
Friday pipeline pass: paste your open threads and have it draft the nudges,
each one referencing the real last conversation.
One post a week from delivery: a pattern you saw (no client specifics) goes in,
a LinkedIn draft comes out, you make it yours and ship it.
The payoff: the pipeline stays warm while you're billing, which is
the whole trick of consulting.
For health advocates
Every client is a crisis, and the paperwork is endless. I'm doing intake,
appeals, and case summaries at 11pm.
The paperwork ally (privacy first)
Tool: Claude Projects or a custom GPT · free tier works
Set the privacy rule first: no client health information in a general AI tool.
Work in templates and blanks; the client's details get added by you, in your own
documents, after.
Create a Project loaded with your service descriptions, your process, and
de-identified skeletons of your best letters.
Build the template library: insurance appeal structures, records-request
letters, provider outreach, new-client onboarding, each with [BRACKETS] where the
person goes.
Draft the public-facing work for real: website copy, workshop outlines, and
the explainers families always need, in plain warm language.
Each new case type becomes a new template. The library compounds; the 11pm
letter becomes a 15-minute fill-in.
The payoff: the paperwork mountain becomes a fill-in-the-blanks
library, and client hours go to clients.
For makers
I make the thing. Then I photograph the thing, list the thing, ship the
thing, and market the thing, and suddenly I never make the thing.
The shop assistant
Tool: Claude Projects or a custom GPT · free tier works
Load your shop: your materials, process, price ranges, three best listings, and
who actually buys from you.
Create a Project that writes listings, wholesale pitches, and market
applications in your voice, specific and unfussy.
Batch the listings: describe each new piece in two plain sentences and get the
full listing back: title, description, materials, care, story.
Draft the wholesale kit once: line sheet copy and a pitch email for the five
shops you'd love to be in, tailored per shop.
Seasonal pushes on tap: holiday collection copy and a posting plan from one
planning chat, while your hands stay on the work.
The payoff: listing night becomes listing hour, and the studio gets
its maker back.
For nonprofit leaders
I'm chasing grants, thanking donors, and writing the newsletter, and the
mission work is what's left over.
The development office of one
Tool: Claude Projects or a custom GPT · free tier works
Load the mission once: your mission and vision, programs, impact numbers,
two funded proposals, and your best appeal letter.
Create a Project instructed to write grants and donor communications that are
concrete and warm, never grandiose.
Grant boilerplate on demand: need statements, program descriptions, and org
history tailored to each funder's language from their actual guidelines.
Donor gratitude that scales: thank-you letters personalized from gift size and
history, drafted for your signature.
The newsletter assembles itself: bring the month's program notes; it drafts,
you add the photo and the human line only you would write.
The payoff: development stops eating the mission's hours, and
funders still hear a human.
For photographers
I'm booked shooting on weekends and editing all week, so inquiries wait
and the blog died in 2023.
The studio front desk
Tool: Claude Projects or a custom GPT · free tier works
Load your studio: packages, pricing, turnaround, FAQs, booking process, and
three inquiry replies that felt right.
Create a Project that answers inquiries warmly, in your voice, with your real
availability rules and next steps.
Every inquiry gets a same-day draft: paste it in, edit the draft, send. Speed
books shoots; you just bought speed.
Automate the client journey copy: booking confirmations, prep guides, gallery
delivery emails, all drafted once, reused forever.
Revive the blog from your galleries: for each delivered shoot, three sentences
from you become an SEO-friendly post it drafts.
The payoff: inquiry response goes same-day without a VA, and
editing week stays for editing.
For realtors
My database is a graveyard. Everyone I've ever sold to is in there, and
none of them have heard from me since closing.
The follow-up machine
Tool: Claude Projects or a custom GPT · free tier works
Load your market: your farm area, your niche, recent sales, and how you talk
to clients (texts included).
Create a Project that writes listing descriptions, nurture emails, and
neighborhood updates that sound like a neighbor, not a brochure.
Resurrect the database: segment past clients into three groups and have it
draft a re-connection note per group that doesn't smell like a campaign.
Monthly value email on rails: feed it three local data points; it drafts the
update people actually read.
Every new listing gets the full kit: MLS description, social captions, open
house follow-up texts, from one paste of the details.
The payoff: the graveyard becomes a referral engine, on one
planning hour a month.
For stylists
I'm behind the chair all day. Rebooking texts, socials, and retail
recommendations all happen on my one day off.
The front-desk-in-your-pocket
Tool: Claude Projects or a custom GPT · free tier works
Load your chair: your services and prices, rebooking windows by service, the
products you actually believe in, and your texting voice.
Create a Project that drafts rebooking nudges, promo messages, and captions
that sound like you between appointments, warm and quick.
Weekly rebooking pass: list who's due ("Maya, balayage, 8 weeks") and get
friendly, specific nudge texts to fire off between clients.
Content from the chair: two sentences about today's transformation becomes a
caption bank entry; batch a month in twenty minutes.
Retail without the pitch: aftercare messages per service that recommend the
product you'd recommend anyway, written once.
The payoff: the day off goes back to being a day off, and the chair
stays booked.
The house rule
Lauren's line, and the standard at this party: what belongs to you and keeps your humanity,
you keep. What frees you to be more human away from your laptop, you hand off. An agent is
just the hand-off getting better. You stay the one who decides what's worth doing.