Agentsphere

A field guide to AI agents. A chatbot answers and waits. An agent takes the next steps for you. Start with the level that matches you. No coding required to begin.

ExplorerStart here if youhave never used an agent, or you only know chatbotsYou will be able toexplain an agent in everyday words, and try one safelyOpen the guide BuilderStart here if youwant to make oneYou will be able tobuild one small agent: a loop, a tool, notes, a skill, and a checkOpen the guide OperatorStart here if yourun agents for a teamYou will be able tochoose where a job runs, and score itOpen the guide Power userStart here if youalready run them, and one is stuckYou will be able tofix it by changing one thingOpen the guide

New here? Open Explorer. It takes about fifteen minutes. You will know what an agent is, and you will have a task you can paste into one.

The rest of the pages live in docs/. Posts and papers that shaped them are in Social references & inspirations. A claim from a post stays linked. A claim from a paper stays linked too.

Choose a level¶

Four levels: Explorer tries one safely, Builder leaves a page, Operator leaves a score, Power user changes one thing.

Level	Start here if you	Open the guide	You will be able to
Explorer	have never used an agent, or you only know chatbots	Explorer	explain an agent in everyday words, and try one safely
Builder	want to make one	Builder	build one small agent: a loop, a tool, notes, a skill, and a check
Operator	run agents for a team	Operator	choose where a job runs, and score it
Power user	already run them, and one is stuck	Power user	fix it by changing one thing

The short sequences and the readiness checks: Learning paths.

If you need to	Open
Explain an agent to someone else	Foundations
See why the same model scores differently	Harness engineering
Leave notes the next chat can find	Wiki memory, then the research playbook
Save a repeated job so you stop pasting it	Write a skill
Choose your laptop or a cloud computer	Operate agents
Decide whether a middleman can read your prompts	Review a gateway
Tell whether a change helped	Score a run
Stop an agent that repeats itself or invents facts	Unstick a loop
Compare Grok Bot, Dots, Muse, and Claude	Hosted agents

The idea under the guide¶

The AI model writes the next step. The system around it decides what the model can see, what it may do, what it remembers, and when it must stop. This guide calls that system the harness. Guo, Hao, Wang, and colleagues lay this out in From Question Answering to Task Completion: A Survey on Agent System and Harness Design. The full page is Harness engineering.

A model inside a six-part harness: observation, context, control, action, state, and verification.

	The model alone	The model plus the harness
What it can do	Suggest a next step in words	See, act, save, and check that step
What is still there tomorrow	Only this chat, if you leave it open	Pages, files, and saved steps
What a score means	"The model scored N"	"This model, in this setup"
Where you change it	A longer prompt	The weak piece: what it sees, what it may do, what it remembers, or the check

The six parts, in one screen

Part	In everyday words	Read
Observation	What it is allowed to look at	The loop
Context	What you put in front of it for this step	How agents remember
Control	Whether it plans, acts, retries, or stops	ReAct, several agents
Action	The move it is allowed to make	Tools
State	What is still there tomorrow	Wiki memory
Verification	The check before the result counts	Threat model

People still mix four habits: a careful prompt, a careful packet of context, a harness around the model, and training that leaves access and safety outside the weights. A later habit does not retire an earlier one.

Map¶

Eleven sections. Each card is a link. The pill is who the section is written for.

Same links, as text:

Foundations — what an agent is, in everyday words.
Agent patterns — how an agent chooses its next step.
Tools — how it connects to apps, files, and the web.
Prompting and skills — instructions you write once and reuse.
Multi-agent — several helpers sharing one job.
Runtime and infra — your laptop, or a cloud computer.
Evals — how you tell the work is actually finished.
Security — what it may do, and what you should check.
Case studies — real projects, told without the hype.
Playbooks — exercises you can follow in one sitting.
Hosted agents — products you can try without installing code.

Also: glossary, how to read a claim, community list, diagram atlas, social references & inspirations. The shelf behind the harness page is Guo et al. 2026.

Products you can rent¶

Some agents are not a repo you clone. They are a product that already chose the computer, the memory, and the gate. The page reads official docs opened on 11 October 2026 and maps each product onto the six harness parts. A feature that was not on those pages is not listed.

Grok Bot, Muse, and Dots keep a cloud computer. Claude and Copilot Studio act on a machine you can point at.

The comparison, the limits, and the power-user checklist are Hosted agents.

How a research agent keeps what it learns¶

An agent remembers only what it writes down outside the context window. The window is working memory: rebuilt every call, and worse at using a fact buried in the middle of a long input than a fact at either end. Episodic memory is the trace of what happened. Semantic memory is a page that says what is true, with a source. Procedural memory is a skill or a checklist. A search that pastes passages into the window is not a fifth kind. It is a load. It becomes memory when the conclusion is written back.

Six moves: name the question, search notes, read sources, write the page, reflect, and let the next run start from that page.

Working memory is the context window. Episodic, semantic, and procedural memory live outside it. Retrieval is loaded for one call unless it is written back.

Memory	Lives	Survives the call?
Working	The window	No
Episodic	Traces, logs, reflections	Yes, if you stored them
Semantic	Wiki pages with sources	Yes, if you update them
Procedural	Skills and checklists	Yes, if they are files

The loop in the first figure is how a research agent uses those stores. Name the decision and the done-when. Search notes and past traces before the open web. Read sources without rewriting them. Write one page: answer, evidence, gap. Reflect as a rule the next run will see, and drop what that rule does not need. The next question starts on the page.

Three methods sit on top of that loop.

Method	The move	Where to read it
Plan, act, observe	Interleave a step with a check of what the tool returned	ReAct
Reflection	Store a judgment that changes the next attempt	Reflexion, Generative Agents
Notes and compaction	Keep a file outside the window. Summarize the window without making the summary the only copy of the decision	Context engineering

Patterns, and what to forget

Five patterns follow from those methods.

Pattern	Use it when
Notes first	The question might already have a page
One page per question	A stranger should be able to answer tomorrow
Claim, source, gap	A number or a product feature is in play
Reflection as a rule	The same miss will happen again
Compact the trace, keep the note	The window is mostly tool output

Forget on purpose. Raw logs can go once the conclusion is on the page. A stale page cannot stay beside a newer one. A reflection nobody will read is just more text in the window.

What you see	Change this
The same search every session	Write the page
A fluent answer with no source	Stop treating retrieval as memory
After a summary, it forgot the task	Put the decision in a note
The same mistake, plus a long self-critique	Store a rule where the next run looks first

Checklist. Done-when is written first. Notes are searched before the web. The page has an answer, a source, and a gap. The raw source was not edited. Something was dropped on purpose. The next run can start without the transcript.

A worked pass. The window is full of test logs and the agent has forgotten the file it was migrating. The note already names billing/client.ts. The run writes the failing assertion into the note, leaves the log on disk, and drops the log body from the window. The next run opens the note. It does not reread the morning's logs to recover a filename.

The full argument, the paper list, and a longer example: How agents remember.

How claims are marked¶

Mark	Meaning
Established	Widely documented practice, a public paper, or an official project page
Per post	Stated in a cited post. Kept, linked, and left unverified by this guide
Synthesis	This guide's own organization of the two

Read the short rules before you quote a number from these pages: How to read agentsphere.

The posts, articles, and repositories behind the pages are listed, with the page each one shaped, in Social references & inspirations.

Extend the guide¶

The writing rules live in CONTRIBUTING. New material is a page a reader will open, with the source linked and the claim marked.