XCrossingCollaborate +
MeaningThe intersection of people and agents
RelationshipAgents with People
StageHuman in the Loop
Core questionHow agents collaborate with people to contribute, correct, and govern knowledge.
PERSONAL RESEARCH NOTES
AgenticXYZ
A coordinate system for Agentic AI.
Building agent-based knowledge collaboration and self-improving agent systems.
More than a name
A coordinate system.
My initials.
A research framework.
AxisMeaningRelationshipStageCore question
YYoursDelegate ×
MeaningAn agent for every person
RelationshipAgents for People
StageHuman on the Loop
Core questionHow agents build personalized memory and continue serving every person.
ZZeroEvolve ^
MeaningTowards self-improvement agentic AI
RelationshipAgents by Agents
StageHuman beyond the Execution Loop
Core questionHow agents improve agents and verify that each update is safe and effective.
A research framework—not a benchmarked maturity scale. The operators are metaphors, not measured performance claims.
X · Crossing
Agents with People.
Humans and agents shape intent, act, verify, and preserve provenance together.
Human in the Loop
Additive leverage1–10×
01Agentic Software
02Knowledge-based PR
03Agentic Runtime
04Agentic Internet
The role of the Harness
Agents should enter environments, take action, and observe consequences.
The Harness connects intent, tools, permissions, memory, state, and outcomes. That is how a model begins to participate in the world.
Y · Yours
Agents for People.
A persistent, user-owned personal agent. You define its goals, permissions, and intervention thresholds; it keeps context and acts over time.
Human on the Loop
Multiplicative leverage10–100×
User-owned memory24 / 7 continuityEscalation by judgmentPortable identity and history
Z · Zero
Agents improving Agents.
Z explores evidence-carrying, gated, and reversible self-improvement—not unconstrained self-modification.
Human beyond the Execution Loop
Compounding leverage100–1000×
Research direction · not a deployed capability claim
The open learning loop
Models now work inside the world.Most of what they experience never becomes verified knowledge.
Agents act, fail, recover, and get corrected. Too often, those outcomes remain logs instead of returning through a structured learning path.
The evolving digital knowledge center
Experience becomes knowledge only when it carries evidence.
01Foundation ModelGeneral capability
→02Harness + SkillsScenarios and methods
→03TrajectoryObservable experience
→04SignalsEvidence and verifiers
Fast updateSkills / MemoryEditable, reversible, immediately reusable
Slow updatePost-trainingOnly stable patterns that generalize
Knowledge moves at three speeds
The model is the capability core—not the only place knowledge lives.
Model parameters
Stable, general, transferable capabilities.
Skills
Explicit, editable, versioned procedural knowledge.
Memory
Private, fresh, and task-specific context.
Four systems · one knowledge lifecycle
From an observed method to a validated capability.
Agentic Software
Discovers methods through human-agent interaction.
Knowledge-based PR
Packages intent, conditions, evidence, and counterexamples.
Agentic Runtime
Tests transfer across tools, schemas, and permissions.
Agentic Internet
Connects capabilities with provenance and authorization.
Post-training candidate
Promotion is considered only after transfer is verified.
Promotion is earned
Zero routine intervention. Never zero human authority.
General?Valid?Novel?Safe?Authorized?Stable?Compatible?
Keep in MemoryUpdate the SkillFreeze as EvaluationPost-training Candidate
Skills do not update model weights by themselves.
Private and time-sensitive knowledge stays external by default.
Internalization requires held-out gains without the original Skill.
Latest from AgenticXYZ
Follow the work as it evolves.
Essays, paper notes, experiments, and moments across Agentic AI.
01
Latest moment · August 14Design Note 02 · August 14, 2026
↗02AgenticXYZ Prototype 1: A Knowledge Collaboration Layer for People and Agents
Manifesto 01 · April 25, 2026
↗AgenticXYZ: Controlled Intelligence, Superintelligent Organizations, and the Fourth Industrial Revolution
GLM-5.3 Released
Z.ai released GLM-5.3, a post-training-driven update focused on stronger coding, longer agent tasks, and newly observed cybersecurity capabilities.
Browse Moments →