← Knowledge Trees · Nestor G Pestelos Jr
AI Agents Knowledge Tree
A structured prerequisite curriculum and mastery map for autonomous AI agent architectures, loops, memory tiers, and multi-agent coordination.
Curriculum Status
22 Concepts organized across 6 learning phases · 14 Mastered · 6 Active Frontier · 2 Planned
Companion Reference: AI Agents Ground-Truth Reference · ELI5: How Do AI Agents Work?
Modeling agent trajectories as POMDPs $(\mathcal{S}, \mathcal{A}, \mathcal{O}, \mathcal{T}, \mathcal{R}, \gamma)$ where LLM token generation parameterizes a policy over discrete tool actions.
Defining rigorous input/output schemas using JSON Schema and Model Context Protocol (MCP) to constrain and parse model tool invocations.
Capturing runtime execution outputs (stdout, stderr, HTTP status, exit codes) and injecting them into the next turn prompt to ground reasoning in physical reality.
Interleaving internal deduction with external tool calls (Thought → Action → Observation → Thought) to avoid speculative error compounding.
Separating high-level strategic roadmap construction from tactical step-by-step execution to maintain long-horizon coherence.
Generating post-mortem linguistic critiques on failed trajectories and persisting reflections into working memory to guide subsequent trials.
Exploring branching decision trajectories using Monte Carlo Tree Search or Breadth/Depth-first search with explicit state rollbacks and pruning heuristics.
Partitioning memory across L1 Working Context (active prompt), L2 Session Scratchpad (ephemeral logs), and L3 Long-term Storage (vector DB/vault).
Preventing chat contamination and token window exhaustion by systematically summarizing completed milestones and dropping raw stdout logs.
Indexing past tool solutions, user preferences, and project facts into dense vector stores with cosine similarity $k$-NN lookups.
Structuring system instructions and tool definitions to maximize KV cache hit rates across multi-turn agent conversations.
Executing agent commands inside ephemeral Docker/WASM sandboxes with restricted network access and strict filesystem boundaries.
Intercepting high-risk system actions (destructive file deletions, migrations, terraform changes) with required human confirmation gates.
Applying unified diffs or contiguous target-replacement blocks with exact collision detection to ensure deterministic codebase modifications.
Automated checkpointing of filesystem and database states prior to multi-file refactors to guarantee clean one-command rollbacks on validation failures.
An orchestrator agent decomposes tasks and dynamically spawns isolated worker subagents with scoped prompts and fresh context windows.
Structuring linear agent workflows where the validated artifact of one specialist becomes the input context for the next stage.
Coordinating asynchronous multi-agent fleets via a shared persistent store (git repository, SQLite DB) without point-to-point message passing.
Running parallel independent reasoning passes with structured cross-critique rounds to filter hallucinations and reach consensus.
Detecting repetitive error-fix cycles and halting execution after $N$ repeated failures to prevent infinite loops and token waste.
Isolating untrusted external data ingestion in a quarantined reader model before passing verified structured summaries to action agents.
Mathematical specification and static assertion checking over agent plans prior to executing production-level mutations.