Find a note that interests you, then explore its prerequisites, references, and related discussions.
234 notes
How execution control, tools, state and verification turn model proposals into bounded work.
How step limits, tool validation, verification, cancellation, and durable state keep agent loops useful and recoverable.
Choose what an AI assistant should remember and retrieve, without confusing stored notes, search results, and authority.
Untrusted content will influence the model, so the execution system must limit what that influence can achieve: what prompt injection, permissions and approvals, sandboxes, credentials, and
Let AI find gaps, explain one step, and provide practice while keeping recall and problem solving with the learner.
Understand production agent harness responsibilities through Codex runtime state, context compaction, integration surfaces, and two-layer security controls.
How to assemble a small, current working set while preserving provenance and durable state outside the model window.
A practical way to label specifications, vendor claims, benchmark results, observations, and recommendations without pretending they prove the same thing.
Recover an agent’s progress after interruption without duplicating the side effects of retries.
A protocol-level view of MCP architecture, lifecycle, primitives, transports, trust boundaries, and alternatives.
When work is split across agents, how to separate context, working directories, and permissions, who integrates and verifies the result, and what an event log can recover.
Understand Pi’s minimal core, customization resources, sessions, and extension boundaries through a documented version.
How to package a repeatable agent procedure, test its trigger and outcome, and manage portability and code-execution risk.
How classical software engineering principles—design concepts, ubiquitous language, TDD, deep modules, and gray-box delegation—resolve failure modes in AI-assisted coding.
Design tools whose inputs, authority, side effects, failures, and evidence remain legible to models and operators.
Block target, group, temporal, and preprocessing leakage by defining prediction time, entities, and train-only pipelines.
Start from deployment distribution, splits, metrics, thresholds, and uncertainty instead of treating one test score as universal ability.
Break requests into measurable stages and explain how concurrency, caching, and batching affect waiting time.
Understand low-bit weight approximation, separate weight, activation, and KV-cache quantization, and estimate deployment costs.
Work through probability correction and timing examples, compare chains, trees, and self-speculation, and connect KV-state correctness to real speed gains.
Distinguish prediction accuracy, reliable probabilities, and worthwhile actions, and use independent data to set calibration and abstention rules.
Measure evidence recall, ranking, faithfulness, and answer quality separately, using counterfactual experiments to locate a RAG system’s bottleneck.
Distinguish model artifacts, evaluation conditions, and real task benefits to make a model-selection decision that can be checked.
Follow next-token distributions to complete answers, including temperature, top-k, top-p, search, stopping, and structured output.
Open an overview to find a topic’s scope and suggested reading routes.