Reference

Python: JSON, types, and dicts

json.loads/dumps, dict key sets, and the action object {tool, args} used everywhere on Joeven.

Python stdlib for agents

The modules Joeven Try-it boxes actually use: json, re, math, hashlib, copy, collections.

Errors, retries, and fail-closed

Return error dicts from tools, retry only timeout/parse/rate_limit, never retry unknown_city.

Vectors, dot product, cosine

The geometry behind RAG: lists of floats, dot product, L2 norm, cosine in [0, 1] for TF vectors.

Probability and sampling

Bernoulli entropy, expected cost, greedy vs weighted sample — decoding in one page.

Softmax and attention weights

Numerically stable softmax, attention as weighted sum, why you subtract max(logits).

Chat messages and roles

system / user / assistant / tool roles, ordering, and what not to put in a system prompt.

Tokens, context, and cost

Rough token math, why loops multiply cost, and what to trim first.

Structured output and JSON schemas

Force {tool, args} or vendor tool_calls; reject extra keys; repair once then fail.

ReAct loop

Thought → Action → Observation until finish; parse text to the same {tool, args} runtime.

Tool schemas and registries

Name → callable, JSON args, error dicts, allowlists. The model never gets globals()[name].

Plan-and-execute vs ReAct

When to write a plan first, when to interleave, and why a supervisor state machine beats a swarm.

Memory, traces, and scratchpads

Transcript vs side trace vs long-term memory. What to store, hash, and drop.

Prompt injection

Untrusted text in pages, tickets, and users. Delimit observations, never execute them, enforce citations and allowlists.

Human approval gates

Approve a digest of tool+args, not a bare id. Mutate never runs on missing/deny/tamper. Stop while waiting.

Sandboxes, allowlists, and exec

Path allowlists, host allowlists, tiny builtins, blocklists as defense-in-depth — not as the only wall.