Interactive lessons - hosted on teach.zahranm.cloud
The architecture of a tool-calling autonomous agent: the conversation loop, tool registry, system prompt assembly, session persistence, gateway pattern, and the five sacred rules.
How Hermes turns one user message into repeated model calls, tool dispatch, recovery, context compression, and a final answer.
How Hermes manages its capability surface across three layers: the tool registry (built-in tools), MCP servers (external tools), and skills (procedural knowledge).
How the system prompt is built from identity, skills, memory, context files, and platform hints.
How conversations are stored, searched, and compressed. The parent_session_id chain pattern.
How one agent core serves 20+ messaging platforms through adapter interfaces.
The self-improvement loop: declarative memory vs procedural skills, and how an agent learns across sessions.
Putting it all together: a working ~200-line Python agent with tools, sessions, and a CLI.
Start from zero: distinguish values, types, functions, actual calls, pairs, and the paper's later symbols. No paper reading required.
Build the first formal model: contexts, reversible effects, inverse composition, effect accumulators, dependency environments, and satisfaction predicates.
Understand sessions, workspaces, tabs, panes, and agent states. Start an agent, detach safely, and reattach without losing work.
Read working, blocked, done, idle, and unknown states; understand detection authority; and diagnose surprising classifications.
Organize multiple agents with workspaces, tabs, panes, Git worktrees, bounded prompts, and safe handoffs.
Choose local, SSH-first, or thin-client remote work; detach safely; and understand restart, native session restore, and live handoff.
Control Herdr from scripts or from inside an agent; choose between layout, pane, agent, and socket primitives.
Measure WPM, accuracy, raw speed, and consistency, then alternate control runs, short ceiling bursts, and real-text transfer.
Try the browser exercise: edit synthetic readings, run a pure calculation, and observe predictable output. No VPS or production data is required.
Use higher-order and polymorphic functions to keep one reusable loop while injecting ordering policy, partial application, and composition.
Model immutable lists with Nil and Cons, follow their shape with pattern matching, share safe suffixes, and generalize recursion with folds.
Holding period returns, expected return, variance, why risk is measured as spread, and the SPY histogram artifact.
Two-asset portfolio variance and the OLS slope derived from first principles; the SPY/IWM scatter artifact.
The utility score, indifference curves, the optimal risky share and minimum-variance weights, and the Sharpe ratio.
PV, FV, compounding, annuities - the financial-math bridge.
The three pillars: faults vs failures, load parameters and fan-out, percentiles, and the three maintainability principles.
From the append-only log to memtables, SSTables, and compaction; the B-tree's in-place pages and WAL; and the write/read trade-off that decides which family your database uses.
Schemas, formats, and how to change data shapes without downtime.
Relational, document, and graph models - and why the object-relational mismatch keeps repeating history.
The whole book on one page: 4 parts, 30 principles, and the three fundamentals (no criticism, honest appreciation, arouse an eager want) that the rest are built on.
Refuse the request, never the person: yes-first, save face, reason in their terms, leave a path — the Steinmetz, Granger, and bow-shop plays.
ASCII diagram mapping all six agentic flows in cv-search.zahranm.cloud: async extraction pipeline (Gemini OCR → pdftotext → LLM → regex fallback), recruiter search, JD matching, job polling, PDF dataset seeding, and CRUD/browse.