AI TEACHING COURSE · LESSON 03

Four theories
of an agent.

OpenClaw · Hermes · Codex · Claude Code — similar capabilities, different engineering philosophies.

~ 45 mindesign comparisonmechanisms, not logos
ML / AI course · Lesson 03← → navigate · S presenter · O overview

THE SAME WORD HIDES DIFFERENT SYSTEMS

An agent product is
a theory of work.

HERMESWork becomes
reusable experience.remember → distill → apply
OPENCLAWWork becomes
personal infrastructure.route → connect → operate
CODEXWork becomes
parallel tasks.delegate → isolate → review
CLAUDE CODEWork becomes
repository context.load → act → hook

DESIGN PHILOSOPHY 01 · HERMES

The agent grows by
turning experience into procedure.

01Conversationtask, correction, recovery
→
02ProposalPROPOSAL.md · evidence · hash
→
03Reviewapprove · reject · revise
→
04Skillreusable procedure next time
Memory stores facts.“This project uses pnpm.”
“The user prefers Chinese notes.”
Skills store procedures.“When a build fails, inspect X, then run Y.”
Self-improvement here means controlled procedural accumulation. It is not silent weight training.

HERMES IS A MEMORY + SKILL RUNTIME

“Learning” is a
filesystem protocol.

~/.hermes/
memories/MEMORY.md
USER.md
skills/agent-created
procedures
cron/scheduled
workflows
sessions/history for
retrieval
ENGINEERING PRINCIPLEKeep continuity cheap, inspectable, and portable across models and surfaces.

A SHARED CONCEPT · SKILLS

A skill is not a new model.
It is a reusable way of working.

EXPERIENCE“Last time, this worked.”correction · recovery · checklist
→
SKILL.mdWhen to useinstructions · tools · examples
progressive disclosure
→
NEXT TASKload only when relevantsame method · new context
Memorystores facts
“this project uses pnpm”
Skillstores procedures
“first inspect X, then run Y”
Modelstores broad capability
language and reasoning patterns
Skill = procedural memory. It changes the agent’s method, not its weights.

DESIGN PHILOSOPHY 02 · OPENCLAW

The agent is
your personal infrastructure.

WhatsApp
Telegram
Slack
iMessage
GATEWAYone agent corerouting · sessions · policy
browser
mac node
skills
workspace
Channel is an adapter, not the identity. OpenClaw makes routing, state, skills, and nodes the stable core.

OPENCLAW DOES NOT EQUATE LEARNING WITH AUTONOMY

The system can learn,
but it learns under governance.

CAPTURE/learnconversation, file, URL, correction
→
PENDINGPROPOSAL.mdscanner · hashes · target binding
→
APPLYlive SKILL.mdrollback metadata before write
30 daysunused skill → stale
90 daysstale skill → archived
pinimportant skills bypass curation
Self-learning ≠ self-authorizing. The proposal queue is part of the product.

A SHARED CONCEPT · MCP

MCP is a protocol
for connecting capability.

HOSTAgent applicationClaude Code · Hermes · another AI app
CLIENTMCP clientone session per server
capability negotiation
SERVERspecialist systemlocal process or remote service
Toolsmodel-callable actions
search · create · update
Resourcescontext and data
files · records · history
Promptsuser-facing templates
repeatable starting points
MCP standardizes the connection. It does not decide whether a tool should be trusted or allowed.

HOW THE PARTS FIT TOGETHER

“How do I work?”
“What can I reach?”
“How do I ship it?”

SKILLMethodinstructions, sequence,
examples, checkshow to work
MCPConnectionserver, tools,
resources, promptswhat to reach
PLUGIN / PACKAGEDistributionskills + apps + config
one installable workflowhow to ship it
EXAMPLE · RESEARCH ASSISTANTSkill: literature-review procedure+MCP: library search + Zotero server+Plugin: package, permissions, and setup

DESIGN PHILOSOPHY 03 · CODEX

Work becomes
parallel, isolated tasks.

mainstable repository
task Afix auth bugagent + sandbox
task Badd testsagent + sandbox
task Creview depsagent + sandbox
REVIEWdiffs · tests · merge
Codex turns autonomy into a queue of reviewable work packages. Parallelism is a product primitive, not just a model trick.

CODEX IS A WORKFLOW SYSTEM

The real output is not just code.
It is evidence.

INPUTTask + repoprompt, files, instructions, skill
RUNSandboxcommands, edits, tests, logs
OUTPUTReview packetdiff, test results, changed files
DECISIONMerge / revise / rejecthuman or policy boundary
ENGINEERING PRINCIPLEMake progress parallel, but make integration reviewable.

DESIGN PHILOSOPHY 04 · CLAUDE CODE

The agent lives inside
the codebase’s local rules.

MANAGED POLICYorganizationsecurity · compliance · shared policy
USER~/.claude/CLAUDE.mdpersonal defaults and preferences
PROJECT./CLAUDE.md · .claude/rules/architecture · commands · team conventions
LOCALCLAUDE.local.mdmachine-specific, personal project context
Context is hierarchical. The closer the rule is to the file being changed, the more local the instruction becomes.

FROM STATIC INSTRUCTIONS TO RUNTIME CONTROL

Claude Code externalizes
the lifecycle.

STARTload rulesCLAUDE.md · memory
→
BEFORE TOOLPreToolUseallow · ask · deny
→
AFTER TOOLPostToolUselint · inspect · log
→
DELEGATEsubagentisolated role + tools
permissionif tool == Bash && command == “rm …” → denyverificationafter file edit → run formatter + test
Claude Code treats process itself as programmable. The agent is embedded in the team’s development lifecycle.

NOW COMPARE THE MECHANISMS

Same capability.
Different center of gravity.

PRODUCTWHAT PERSISTS?WHAT IS PARALLEL?WHERE IS CONTROL?
Hermesfacts + proceduresdelegated subagentsmemory / skill approval
OpenClawpersonal state + channelsagents / nodes / channelsgateway + workshop
Codexskills + repo workflowworktrees / cloud taskssandbox + review
Claude Coderules + project memorysubagents / teamspermissions + hooks
The product difference is visible in the answer to one question: what does the system remember when the current turn ends?

THE SAME REQUEST REVEALS THE PHILOSOPHY

The verb “automate”
means four different things.

Hermeslearn the briefing procedurememory + skill + cron + delivery
OpenClawreach me wherever I amgateway + channel binding + workspace
Codexmake the artifact reviewablerepo + task run + diff + tests
Claude Coderespect the project’s rulesCLAUDE.md + hooks + permissions

A NECESSARY DISTINCTION

Agent evolution is usually
around the model.

WEIGHT UPDATEtraining / fine-tuningchanges model parameters
expensive · broad · harder to audit
≠
SYSTEM UPDATEmemory / skill / rule / toolchanges context or procedure
local · reversible · task-specific
Hermesdistill experience
OpenClawpropose + curate skills
Codexsave repeatable workflows
Claude Codeversion rules + hooks
Question for students: Which layer changed when the agent “got better”?

WHICH PHILOSOPHY FITS YOUR TASK?

Choose the center
of gravity you need.

What must stay stable?
Personal historyChoose a memory + skill runtime.→ Hermes
Many channels and devicesChoose a personal control plane.→ OpenClaw
Parallel engineering outputChoose isolated tasks and review packets.→ Codex
Repository rules and lifecycleChoose local context, hooks, and permissions.→ Claude Code

CLASSROOM LAB · 20 MINUTES

Reverse-engineer one product
from its mechanisms.

01StateWhat survives the turn?
02ImprovementHow does experience become reusable?
03ParallelismWhat can run at the same time?
04ControlWhere can a human stop it?
TEAM PROMPTChoose Hermes, OpenClaw, Codex, or Claude Code. Draw its persistence path, action boundary, and failure recovery path. Then name one design trade-off.
GOOD ANSWER“This product is optimized for X, because mechanism Y makes Z persistent—but that creates risk W.”

TAKEAWAY

Different agents
make different things stable.

HERMESexperiencefacts → procedures
OPENCLAWinfrastructurechannels → gateway
CODEXthroughputtasks → review packets
CLAUDE CODEpracticerules → lifecycle
Next lesson: build a minimal learning loop, then decide what should become memory, skill, rule, or test.

SOURCES · ACCESSED 2026-07-30

The details are
in the protocols.

Hermes · Skill Workshopproposal → review → apply · procedural memoryOpenClaw · Self-learningpending proposals · no silent live-skill writesOpenAI · Codexworktrees · cloud environments · parallel agents · skillsClaude Code · Memoryhierarchical CLAUDE.md context and project rulesClaude Code · Hooksruntime lifecycle and tool interceptionMCP · ArchitectureHost · Client · Server · capability negotiation