CS2680 Lecture 8 (2026-10-07) has two halves. The first takes apart an agent's bill: cost is a product of users × sessions per user × turns per session × steps per turn × tokens per step × price per token, and the last four are the ones you can cut. The students' own Claude Code logs show 97.6% of tokens are cached re-reads, so keeping the prefix stable matters more than keeping the prompt short. Speed is set by the critical path, not the sum of every step. The second half compares three harness architectures (monolithic, modular, microkernel) and the design choices of eight open-source and commercial harnesses. This post works one bill by hand: put a timestamp at the start of the system prompt and the same task goes from 23.6 units back to 65 or more.
On 1M-context models, Claude Code waits until about 967K tokens to auto-compact, so late messages each carry nearly a million tokens of history. I set /autocompact 400k as Lydia Hallie suggested and used it for 11 days: roughly 16–22% less usage, and no noticeable increase in rework or corrections from lost details. This is one person's observation on one kind of work, not a general finding.
REA is an MIT-licensed MCP server that lets Claude Code, Codex, Cursor, and other agents read native binaries through Ghidra, Hopper, or IDA, and take apart Electron, .NET, APK, and web targets directly. What separates it from single-engine bridges like GhidraMCP is that every result carries an evidence ID, a coverage status, and its limitations. One npx rea-agents setup installs it, and JavaScript analysis needs no disassembler at all.
A Mac has no NVIDIA GPU, so it can't really run vLLM — but Colab's free T4 (compute capability 7.5, exactly vLLM's minimum) can. Serve the 0.6B Qwen3 as an OpenAI-compatible API with vllm serve, fire 8 requests at once, read /metrics, and you can watch continuous batching happen.
To cut the input tokens resent every turn, agent platforms often truncate, clear, or summarize old tool results, and skill instructions, write receipts, and error messages get cut along with them. The Complexity Trap shows observation masking can halve cost, yet Gemini 2.5 Flash with thinking still dropped from 40.4% to 36.4% solve rate. Measure savings against full history and fixes against the current setup, and judge by total cost per successful task, not shorter per-turn input.
Claude's docs say tool-selection accuracy degrades past 30–50 tools, and both Claude and OpenAI now load tools on demand via tool search. But finding a tool isn't finishing the task: on ToolRet's 43k-tool benchmark, no retriever put the complete required toolset in its top 10 more than 35% of the time. This post splits the problem into five failure points, maps seven approaches onto them, and lays out an evaluation.
Anthropic engineering lead Felix Rieseberg splits 'being more ambitious' into Speed, Altitude, Depth, and Range, then uses one quarterly business review (QBR) to show how to prompt and verify for each of the last three.
Matt Pocock posted 18 'mistakes you might be making' on X with no explanations. Grouped against his public skills and talks, they fall into five: verification, letting go, context, alignment, thinking. Six map directly onto /retro's check categories, so you can try one session tonight.
pstack is an MIT-licensed set of agent skills from Cursor engineer Lauren Tan. One entry point, /poteto-mode, picks one of 23 playbooks per task and calls other skills and 24 principles as steps need them. Install takes one command and one model setup; the price is extra tokens for subagents and review panels.
A chat model is rewarded for a single response people prefer; an agent has to be rewarded for whether a multi-step task actually finished, which needs different data and environments.
Reasoning is the intermediate tokens between input and output; RL fine-tuning beats imitating human solutions because of a reliable verifier, and sampling several times then taking the most frequent answer pushes accuracy further.
AlphaGeometry2 relies on a symbolic engine and AlphaProof on Lean, for a combined 28 points in 2024; in 2025 Gemini Deep Think scored 35 points in natural language, with verification done by human graders.
Laskin's view is that engineers spend 80% of their time understanding systems rather than writing code, and that how far RL goes depends on clean rewards, a problem he considers as hard as superintelligence itself.
A VLA adds an action module to a vision-language model; letting that module's gradients flow back harms language understanding, so knowledge insulation blocks the gradient, while π0.5 co-trains on web data and data from many robots to gain generalization.
Alexandr Wang says Meta listed 100+ behaviors a personal agent needs, built an eval for each, and set two thresholds per row: a launch-blocking floor and a bar for 'good'. They shipped after the first all-green checkpoint and a hands-on check that it was clearly better; they also ran new-user testing. The method is reusable, but post-launch safety reports show all-green is not the same as safe.
This lecture puts the first concrete content into the flowchart from Lecture 1: the hypothesis set H is all lines (perceptrons), and the algorithm A is PLA, which adds yₙxₙ to w whenever it sees a misclassified point. When the data is linearly separable, PLA started from w₀ = 0 halts after at most R²/ρ² corrections; when the data cannot be separated, switch to pocket, which keeps the line with the fewest mistakes seen so far in its pocket.
This lecture takes the learning flow diagram from Lecture 1 and swaps one part at a time along four directions: the output space Y (binary classification, multiclass classification, regression, structured learning), the labels yₙ (supervised, unsupervised, semi-supervised, reinforcement learning), the way data reaches the algorithm (batch, online, active), and the input space X (concrete, raw, abstract features). Each direction has one core option, and together they mark the home ground for the rest of Foundations: batch, supervised binary classification and regression on concrete features.
Once a hypothesis set has a break point k, the largest number of dichotomies it can produce on N points, B(N, k), satisfies B(N, k) ≤ B(N−1, k) + B(N−1, k−1). That gives B(N, k) ≤ C(N, 0) + … + C(N, k−1), whose highest term is N to the power k−1. Three more steps put the growth function into Hoeffding and yield the VC bound: the probability of the bad event is at most 4·m_H(2N)·exp(−ε²N/8). The break point of 2D perceptrons is 4, so they can learn.
This lecture replaces the fixed target function with a target distribution P(y|x): as long as each (x, y) is drawn independently from the same joint distribution, the VC bound holds with noise just as before. The error measure err and P(y|x) together decide the ideal prediction on each x; the 0/1 error picks the most probable y and the squared error picks the weighted mean. CIA fingerprint verification penalizes a false accept 1000 times more, and this kind of weighted classification is the same as copying the −1 examples 1000 times and then doing ordinary classification.
Logistic regression uses θ(wᵀx) to estimate P(+1|x), yet the data contains only circles and crosses. This lecture derives the cross-entropy error ln(1 + exp(−y wᵀx)) from maximum likelihood and uses the chain rule to show that the gradient is a θ-weighted average of −yₙxₙ. Setting the gradient to zero gives a non-linear equation of w with no closed-form solution, so the lecture switches to fixed learning rate gradient descent, w ← w − η∇E_in(w), and walks to the bottom of the valley one step at a time.