-
Multi-Agent Systems Explained: A Sub-Agent Is Just a Tool That Runs Its Own Loop (Ep:06.08)
An orchestrator agent invoking a specialist sub-agent is structurally identical to a single agent calling a tool — the only difference is that “executing” the action means running an entire…
-
Why “More Training” Didn’t Fix the Undertrained Tool — a Deeper Diagnosis (Ep:06.07)
Oversampling a tool’s training data 4x more didn’t fix it — because the real problem wasn’t repetition count, it was the size of the tool’s input domain relative to what…
-
Why Self-Correction (Reflexion) Doesn’t Fix an Undertrained Model (Ep:06.06)
Reflexion lets an agent verbally reflect on its own mistakes and retry, entirely in-context, with no weight updates. Module 6 of From Zero to Agents tests it directly on a…
-
Fixing a Broken Evaluation Metric: Comparable Positions, Not Comparable Sequences (Ep:06.03)
The fix for a per-tool evaluation metric confounded by trace length isn’t more data — it’s evaluating only the one position that’s genuinely comparable across every tool. Module 6 of…
-
Catching an Undertrained Tool Before It Fails: Two Signals, and a Real Evaluation Trap (Ep:06.02)
Per-tool held-out loss and label-free generation confidence both aim to catch an undertrained agent tool automatically — but one of them walks straight into the exact evaluation-scope trap Module 5…
-
Multi-Tool Agents and Chained Tool Calls: Extending the ReAct Loop (Ep:06.01)
Choosing between tools is just next-token prediction over a bigger action vocabulary, and chaining tool calls is the same loop run more than once. Module 6 of From Zero to…
-
Building a Tool-Using Agent From Scratch: The ReAct Loop Explained (Ep:06.00)
An agent isn’t a new model architecture — it’s a loop, built entirely on top of the causal transformer from Module 4, that lets a model call real tools and…
-
A Complete LoRA Fine-Tuning Pipeline, End to End: Every Module 5 Concept in One Run (Ep:05.08)
Pretrain a base model on one task, freeze it, LoRA-adapt it to a second task with proper loss masking, then measure new-task accuracy, old-task interference, and exact recoverability — every…
-
Perplexity Can Lie: Why Which Tokens You Evaluate On Changes the Answer (Ep:05.07)
Full-sequence perplexity makes an unmasked-trained model look better than a masked-trained one — the opposite of the truth. Module 5 of From Zero to Agents derives perplexity from cross-entropy, then…
-
Loss Masking, Properly Tested: Why the First Experiment Failed and the Second One Worked (Ep:05.06)
A properly designed experiment — more data, a task requiring genuine rule-learning, a real held-out test set — finally shows loss masking’s real benefit clearly, and reveals a precise mechanistic…
