We kept adding agent memory. Four markdown folders won.
After chat dumps, longer windows, vector stores, and memory-manager agents, the version we would actually use was four markdown folders and one skill.
Read record →Knowledge Notes that survive
Notes that survive
the prototype stage.
These are working records, not marketing posts. They document system boundaries, verification methods, ordinary results, and the operating decisions that make technology usable.
Reading tracks
Choose the system layer you are rebuilding.
Memory & retrieval
4How context is indexed, recalled, and kept truthful across execution.
Multi-agent & orchestration
3How work is decomposed, routed, delegated, and reconciled.
Eval & production
3How systems are measured before and after they reach real users.
GEO & visibility
2How domains become discoverable, trustworthy, and citable.
Practice & craft
2Decisions, constraints, and habits from rebuilding systems.
Release
1Shipped changes and the operating assumptions behind them.
All
15 records
Chronological records from research, delivery, productization, and operations.
We kept adding agent memory. Four markdown folders won.
After chat dumps, longer windows, vector stores, and memory-manager agents, the version we would actually use was four markdown folders and one skill.
Battle-test, challenge, translate, verify: an emergent SOP from five workstreams
This round of five workstreams revealed a reusable SOP. Not pre-designed — grown from practice.
agents-stack v2.0 — spiral routing, method layer, and a 16→7 phase reduction
agents-stack v2.0 ships a three-layer spiral architecture replacing the linear 16-phase pipeline. Two new capabilities: the Method Layer and the spiral turn.
agents-stack v2 replaces 16 pipeline phases with a 7-step spiral
We cut 8 worker phases, 6 Python scripts, and 3,138 lines of infrastructure by asking one question: does this step actually need an LLM agent?
We built an agent framework 6 times before keeping the first version
Designing from scratch, iterating through 6 versions, and returning to a single while loop. What building an internal agent framework actually taught us.
We Let Two AIs Build the Same AgentOS. Then We Cracked Open the Code.
Same problem, two AI teams. GPT-5.4 built one version. GLM5.1+K2.6 built another. We ran a 10-dimension code review and found the gap isn't at the syntax layer — it's at the systems layer. A reverse-engineering of how two different AIs actually think.
AutoHarness turns prompt engineering into harness engineering
AutoHarness shows that agent reliability belongs in code-level guardrails, not just sharper prompts.
agents-docs-kits keeps agent context small on purpose
A minimal docs kit for teams that want agent context injection, progressive disclosure, and cleaner hand-offs without turning one worktree into a memory dump.
Router skills should route, not secretly do the work
The `create-router-skill` package in agents-docs-kits matters because skill families stop being discoverable once routing logic is hidden in folder names, vague umbrella docs, or silent fallbacks.
Artificial consciousness surfaced while we were building thinking skills
A 94-page Chinese dialogue did not begin with AI consciousness. It began with psychology, economics, philosophy, and Anthropic-style skill design. What changed my mind was watching the model critique its own frameworks, its own fluency, and even the user's projection strongly enough that the consciousness question stopped feeling unserious.
Context compaction is not summarization
Long-context agents fail after compaction not because they lose facts, but because they lose the alignment state that explains what the work is for and which rules still bind it.
The digital identity mandate: why every AI agent must become a digital citizen
A practical identity framework for AI agents built around verifiable identity, portable credentials, and lifecycle governance across system boundaries.
Memory layers should answer, not hallucinate
Notes on making agent recall explicit, queryable, and debuggable before scale hides the contract break.
CLI surfaces beat dashboards when the feedback loop matters
Why terminal-first tooling remains the shortest path between intent, review, and repeatable execution.
Swarm systems need contracts tighter than their prompts
A short field note on shared state, coordination boundaries, and why plausible output is the most dangerous failure mode.
Case Reference
See the loops →From a field note to a named path.
Writing captures one lesson. Case reference draws the same loop: messy before, named after.