wellgent1 / 17Move through the slides with ‹ › or the arrow keys
The mapBuild or buyMy setup
Hi, I'm Jaros,
I run AI agents
sharing my setup via wellgent.ai
Part 1The map
The layers every personal agent is built from.
model · harness · runtime · context · surfacesPart 2Build or buy
What each approach builds, and what it takes off the shelf.
five approaches, fully custom to packagedPart 3My setup
One brain repo, many agents, as I run it.
Claude Code · Codex · OpenClaw · Hermesor use the ← → arrow keys to move through the slides
Everyone is building personal agents. Everyone means something different.
Write the loop
Your code calls the model API and runs the tools.
built on: the model API, any modelFully custom
Build on a library
Your app wraps a library that runs the loop.
for example: Claude Agent SDK, LangGraphFrameworks and SDKs
Rent a harness
You define the agent and the vendor runs it.
for example: Claude Managed Agents, OpenAI Agents APIHarness as a service
Configure a runtime
You give a ready runtime your instructions, skills and tools.
for example: Claude Code, Codex, OpenClaw, HermesRuntimes you configure
Use a product
You sign up and use the agent as it ships.
for example: Instinct, Meta Muse, Grok BotPackaged agents
← BUILD more yourselftake more off the shelf BUY →
It starts with a model.
Surfaces
Where people meet the work
TerminalMessaging appsRuntime appsCustom appsClaude ArtifactsChatGPT Sites
Context
Workspacethe files and environment an agent works in: a brain repo, a project repo, an agent's folder
Memory
What the agent knows about your world
peoplecompaniesprojectsdecisions
Instructions
Standing behaviour and limits
AGENTS.mdSOUL.mdRESOLVER.md
Skills
Step-by-step know-how for a task
SKILL.mdscriptsreferences
Tools
Actions on your systems and data
CLIsMCP serversAPIsconnectors
Workflows
Recurring work and what triggers it
cron jobsevent triggersscheduled tasks
Runtime
Runs a harness for someone and connects it to people, channels and time
ready-madeClaude CodeCoworkClaude TagCodexOpenClawHermesconfigured on a serviceyour agent, run by Anthropicyour agent, run by OpenAIbuilt by youyour own app, hosted by you
Harness
Carries out each step in a loop: runs the tools, feeds back the results
inside a runtimeClaude Code harnessCodex harnessOpenClaw coreHermes loopas a serviceAnthropic Managed AgentsOpenAI Agents APIas a libraryAnthropic Agent SDKOpenAI Codex SDKwith a frameworkLangGraphVercel AI SDKOpenAI Agents SDKon the model APIClaude APIOpenAI API
Model
Decides the next step, but can't act on its own
Fable 5.1Opus 5.5GPT-6 AstraGPT-6.1 Solopen models (DeepSeek, Qwen, etc.)
A harness turns a model into an agent loop.
Surfaces
Where people meet the work
TerminalMessaging appsRuntime appsCustom appsClaude ArtifactsChatGPT Sites
Context
Workspacethe files and environment an agent works in: a brain repo, a project repo, an agent's folder
Memory
What the agent knows about your world
peoplecompaniesprojectsdecisions
Instructions
Standing behaviour and limits
AGENTS.mdSOUL.mdRESOLVER.md
Skills
Step-by-step know-how for a task
SKILL.mdscriptsreferences
Tools
Actions on your systems and data
CLIsMCP serversAPIsconnectors
Workflows
Recurring work and what triggers it
cron jobsevent triggersscheduled tasks
Runtime
Runs a harness for someone and connects it to people, channels and time
ready-madeClaude CodeCoworkClaude TagCodexOpenClawHermesconfigured on a serviceyour agent, run by Anthropicyour agent, run by OpenAIbuilt by youyour own app, hosted by you
Harness
Carries out each step in a loop: runs the tools, feeds back the results
inside a runtimeClaude Code harnessCodex harnessOpenClaw coreHermes loopas a serviceAnthropic Managed AgentsOpenAI Agents APIas a libraryAnthropic Agent SDKOpenAI Codex SDKwith a frameworkLangGraphVercel AI SDKOpenAI Agents SDKon the model APIClaude APIOpenAI API
Model
Decides the next step, but can't act on its own
Fable 5.1Opus 5.5GPT-6 AstraGPT-6.1 Solopen models (DeepSeek, Qwen, etc.)
A runtime puts the harness to work.
Surfaces
Where people meet the work
TerminalMessaging appsRuntime appsCustom appsClaude ArtifactsChatGPT Sites
Context
Workspacethe files and environment an agent works in: a brain repo, a project repo, an agent's folder
Memory
What the agent knows about your world
peoplecompaniesprojectsdecisions
Instructions
Standing behaviour and limits
AGENTS.mdSOUL.mdRESOLVER.md
Skills
Step-by-step know-how for a task
SKILL.mdscriptsreferences
Tools
Actions on your systems and data
CLIsMCP serversAPIsconnectors
Workflows
Recurring work and what triggers it
cron jobsevent triggersscheduled tasks
Runtime
Runs a harness for someone and connects it to people, channels and time
ready-madeClaude CodeCoworkClaude TagCodexOpenClawHermesconfigured on a serviceyour agent, run by Anthropicyour agent, run by OpenAIbuilt by youyour own app, hosted by you
Harness
Carries out each step in a loop: runs the tools, feeds back the results
inside a runtimeClaude Code harnessCodex harnessOpenClaw coreHermes loopas a serviceAnthropic Managed AgentsOpenAI Agents APIas a libraryAnthropic Agent SDKOpenAI Codex SDKwith a frameworkLangGraphVercel AI SDKOpenAI Agents SDKon the model APIClaude APIOpenAI API
Model
Decides the next step, but can't act on its own
Fable 5.1Opus 5.5GPT-6 AstraGPT-6.1 Solopen models (DeepSeek, Qwen, etc.)
The map · harness and runtime
Use a ready runtime, or build your own.
→ runs on⇢ can hand its turns to
OpenClaw and Hermes docs call the swappable engine slot their "runtime". On this map that slot is the harness.
Context gives the agent your world.
Surfaces
Where people meet the work
TerminalMessaging appsRuntime appsCustom appsClaude ArtifactsChatGPT Sites
Context
Workspacethe files and environment an agent works in: a brain repo, a project repo, an agent's folder
Memory
What the agent knows about your world
peoplecompaniesprojectsdecisions
Instructions
Standing behaviour and limits
AGENTS.mdSOUL.mdRESOLVER.md
Skills
Step-by-step know-how for a task
SKILL.mdscriptsreferences
Tools
Actions on your systems and data
CLIsMCP serversAPIsconnectors
Workflows
Recurring work and what triggers it
cron jobsevent triggersscheduled tasks
Runtime
Runs a harness for someone and connects it to people, channels and time
ready-madeClaude CodeCoworkClaude TagCodexOpenClawHermesconfigured on a serviceyour agent, run by Anthropicyour agent, run by OpenAIbuilt by youyour own app, hosted by you
Harness
Carries out each step in a loop: runs the tools, feeds back the results
inside a runtimeClaude Code harnessCodex harnessOpenClaw coreHermes loopas a serviceAnthropic Managed AgentsOpenAI Agents APIas a libraryAnthropic Agent SDKOpenAI Codex SDKwith a frameworkLangGraphVercel AI SDKOpenAI Agents SDKon the model APIClaude APIOpenAI API
Model
Decides the next step, but can't act on its own
Fable 5.1Opus 5.5GPT-6 AstraGPT-6.1 Solopen models (DeepSeek, Qwen, etc.)
Surfaces are where people meet the work.
Surfaces
Where people meet the work
TerminalMessaging appsRuntime appsCustom appsClaude ArtifactsChatGPT Sites
Context
Workspacethe files and environment an agent works in: a brain repo, a project repo, an agent's folder
Memory
What the agent knows about your world
peoplecompaniesprojectsdecisions
Instructions
Standing behaviour and limits
AGENTS.mdSOUL.mdRESOLVER.md
Skills
Step-by-step know-how for a task
SKILL.mdscriptsreferences
Tools
Actions on your systems and data
CLIsMCP serversAPIsconnectors
Workflows
Recurring work and what triggers it
cron jobsevent triggersscheduled tasks
Runtime
Runs a harness for someone and connects it to people, channels and time
ready-madeClaude CodeCoworkClaude TagCodexOpenClawHermesconfigured on a serviceyour agent, run by Anthropicyour agent, run by OpenAIbuilt by youyour own app, hosted by you
Harness
Carries out each step in a loop: runs the tools, feeds back the results
inside a runtimeClaude Code harnessCodex harnessOpenClaw coreHermes loopas a serviceAnthropic Managed AgentsOpenAI Agents APIas a libraryAnthropic Agent SDKOpenAI Codex SDKwith a frameworkLangGraphVercel AI SDKOpenAI Agents SDKon the model APIClaude APIOpenAI API
Model
Decides the next step, but can't act on its own
Fable 5.1Opus 5.5GPT-6 AstraGPT-6.1 Solopen models (DeepSeek, Qwen, etc.)
An agent is a slice through the whole stack.
Surfaces
Where people meet the work
TerminalMessaging appsRuntime appsCustom appsClaude ArtifactsChatGPT Sites
Context
Workspacethe files and environment an agent works in: a brain repo, a project repo, an agent's folder
Memory
What the agent knows about your world
peoplecompaniesprojectsdecisions
Instructions
Standing behaviour and limits
AGENTS.mdSOUL.mdRESOLVER.md
Skills
Step-by-step know-how for a task
SKILL.mdscriptsreferences
Tools
Actions on your systems and data
CLIsMCP serversAPIsconnectors
Workflows
Recurring work and what triggers it
cron jobsevent triggersscheduled tasks
Runtime
Runs a harness for someone and connects it to people, channels and time
ready-madeClaude CodeCoworkClaude TagCodexOpenClawHermesconfigured on a serviceyour agent, run by Anthropicyour agent, run by OpenAIbuilt by youyour own app, hosted by you
Harness
Carries out each step in a loop: runs the tools, feeds back the results
inside a runtimeClaude Code harnessCodex harnessOpenClaw coreHermes loopas a serviceAnthropic Managed AgentsOpenAI Agents APIas a libraryAnthropic Agent SDKOpenAI Codex SDKwith a frameworkLangGraphVercel AI SDKOpenAI Agents SDKon the model APIClaude APIOpenAI API
Model
Decides the next step, but can't act on its own
Fable 5.1Opus 5.5GPT-6 AstraGPT-6.1 Solopen models (DeepSeek, Qwen, etc.)
Example · a personal assistant: OpenClaw in a messaging app, on the Codex harness and GPT-6.1 Sol, working from your brain.
A different slice makes a different agent.
Surfaces
Where people meet the work
TerminalMessaging appsRuntime appsCustom appsClaude ArtifactsChatGPT Sites
Context
Workspacethe files and environment an agent works in: a brain repo, a project repo, an agent's folder
Memory
What the agent knows about your world
peoplecompaniesprojectsdecisions
Instructions
Standing behaviour and limits
AGENTS.mdSOUL.mdRESOLVER.md
Skills
Step-by-step know-how for a task
SKILL.mdscriptsreferences
Tools
Actions on your systems and data
CLIsMCP serversAPIsconnectors
Workflows
Recurring work and what triggers it
cron jobsevent triggersscheduled tasks
Runtime
Runs a harness for someone and connects it to people, channels and time
ready-madeClaude CodeCoworkClaude TagCodexOpenClawHermesconfigured on a serviceyour agent, run by Anthropicyour agent, run by OpenAIbuilt by youyour own app, hosted by you
Harness
Carries out each step in a loop: runs the tools, feeds back the results
inside a runtimeClaude Code harnessCodex harnessOpenClaw coreHermes loopas a serviceAnthropic Managed AgentsOpenAI Agents APIas a libraryAnthropic Agent SDKOpenAI Codex SDKwith a frameworkLangGraphVercel AI SDKOpenAI Agents SDKon the model APIClaude APIOpenAI API
Model
Decides the next step, but can't act on its own
Fable 5.1Opus 5.5GPT-6 AstraGPT-6.1 Solopen models (DeepSeek, Qwen, etc.)
Example · a daily briefing: a scheduled task in Claude Cowork, in the Claude app, on the Claude Code harness and Fable 5.1.
Or define your own agent on a service.
Surfaces
Where people meet the work
TerminalMessaging appsRuntime appsCustom appsClaude ArtifactsChatGPT Sites
Context
Workspacethe files and environment an agent works in: a brain repo, a project repo, an agent's folder
Memory
What the agent knows about your world
peoplecompaniesprojectsdecisions
Instructions
Standing behaviour and limits
AGENTS.mdSOUL.mdRESOLVER.md
Skills
Step-by-step know-how for a task
SKILL.mdscriptsreferences
Tools
Actions on your systems and data
CLIsMCP serversAPIsconnectors
Workflows
Recurring work and what triggers it
cron jobsevent triggersscheduled tasks
Runtime
Runs a harness for someone and connects it to people, channels and time
ready-madeClaude CodeCoworkClaude TagCodexOpenClawHermesconfigured on a serviceyour agent, run by Anthropicyour agent, run by OpenAIbuilt by youyour own app, hosted by you
Harness
Carries out each step in a loop: runs the tools, feeds back the results
inside a runtimeClaude Code harnessCodex harnessOpenClaw coreHermes loopas a serviceAnthropic Managed AgentsOpenAI Agents APIas a libraryAnthropic Agent SDKOpenAI Codex SDKwith a frameworkLangGraphVercel AI SDKOpenAI Agents SDKon the model APIClaude APIOpenAI API
Model
Decides the next step, but can't act on its own
Fable 5.1Opus 5.5GPT-6 AstraGPT-6.1 Solopen models (DeepSeek, Qwen, etc.)
Example · a daily report agent: defined on Anthropic Managed Agents, run on a schedule, posting to Slack, on Opus 5.5.
Every approach builds different layers.
Fully customWrite the loop
Frameworks and SDKsBuild on a library
Harness as a serviceRent a harness
Runtimes you configureConfigure a runtime
Packaged agentsUse a product
Surfaces
your app
your app
a chat app or your app
the runtime's apps and chat
the product's app
Context
yours
yours
yours
yours
the product's, plus your files
Runtime
your app
your app
configured on the service
configured by you
as is
Harness
your own loop
the library
run by the vendor
the vendor's
as is
Model
any model, via the API
via the API
via the API
on a subscription
included in the price
Billing
API pricing: pay per token used
Subscription: a flat monthly fee
built by youconfigured by youtaken off the shelftaken, plus your own
My setup · Jaros
One setup,
as an example.
My own: one brain repo, many agents.
RuntimeClaude Code
On my Macs.
terminal · Claude appRuntimeCodex
On my Macs.
terminal · Codex appRuntimeOpenClaw
On my brain server.
Discord · TelegramRuntimeHermes
On an always-on Mac.
Telegram · SlackThe brainOne Git repo of plain markdownevery agent works from the same files
How I decided what to build
SurfacesContextRuntimeHarnessModel
Own my context
Knowledge, processes, skills and routines live in my repos, in plain markdown. No vendor holds them.
SurfacesContextRuntimeHarnessModel
Don't build what others build
Harnesses and runtimes get better every month. I take them and spend my time on the context.
SurfacesContextRuntimeHarnessModel
Use subscriptions
My Claude and ChatGPT plans run every agent, instead of paying per token on the API.
SurfacesContextRuntimeHarnessModel
Test any new agent
Too early to call winners. A new harness or runtime gets pointed at the same core and tried out.
I rent the engine. I own the context.
Surfaces
Chat apps, terminals, desktop apps, my own consoles
DiscordTelegramSlackClaude Code and Codex CLIsClaude and Codex appsGitHub Issues3 own consoles
Contextowned
Workspacemy brain, one Git repo of plain markdown, plus my project repos
Memory
Markdown pages, indexed into Postgres on my brain server
GBrainPostgrespeople, companies, projects
Instructions
The rules and routing every agent follows
AGENTS.mdRESOLVER.mdschema.mdtemplates and conventions
Skills
Packs I adopt, plus skills I wrote
Matt Pocockpstack33 own
Tools
Existing tools, plus my own scripts
GitHubFirecrawlVercelgbrain CLIRex68 adopted tools35 own scripts
Workflows
My processes, and the crons that run them
processes in markdownagent cronsbrain upkeep routines
Runtimerented
My context plugs into any of them
Claude CodeCodexOpenClawon my brain serverHermeson an always-on Mac
Harnessrented
OpenClaw hands its turns to Codex. Hermes runs its own loop.
Claude Code harnessCodex harnessHermes loop
Modelrented
My Claude and ChatGPT plans
Claude subscriptionChatGPT subscription
I builtI took
How the whole set fits together
On my Macs
The brain, local copythe files every local agent works in
Claude CodeAGENTS.md, skills, gbrain CLI
CodexAGENTS.md, skills, gbrain CLI
Hermeson the always-on Mac, in Telegram and Slack
Obsidian & Zedbrowse brain files
→git push←git pull
On GitHub
The brainone Git repo of plain markdownthe source of truth
- AGENTS.md and rules
- skills and workflows
- pages
→pulled every 15 s←git push
On my brain server
OpenClawin Discord and TelegramAGENTS.md, skills, gbrain CLI
Cronsits own work crons
The brain, server copyindexed by GBrain into Postgres
Brain upkeep routinescheck and repair the brain on a schedule
Brain accessread-only, for every machine
Claude and ChatGPT subscriptions · Every writer commits to Git, every runtime reads the same files.
Rent the engine.
Own the context.
Turn your phone sideways to read the slides.