Stanley Yang — Writing

Create concise repository instructions that expose non-obvious commands, boundaries, architecture, and verification without duplicating discoverable documentation.

Implement escalation rules for ambiguity, missing authority, weak evidence, uncertain side effects, and high-impact decisions.

Evolve tools, resources, prompts, transports, and experimental features using protocol negotiation, additive contracts, compatibility tests, and staged removal.

Preserve agent-session UI state across tabs while correctly stopping effects, controlling memory, and protecting sensitive sessions.

Design time-ordered, idempotent agent-event tables with PostgreSQL 18 uuidv7 without confusing identity, time, and sequence.

Have coding agents plan and verify expand-migrate-contract changes without granting them unsupervised production database access.

Create versioned plans, reports, patches, and decisions that are validated at runtime instead of trusting model-generated prose.

Design and evaluate voice-agent turn taking without treating every silence as completion or every sound as an interruption.

Use an exact-search baseline and a repeatable benchmark to tune pgvector HNSW construction, query, filtering, and memory parameters.

Evaluate agent tool selection, arguments, ordering, evidence, side effects, and efficiency without requiring one brittle golden path.

Instrument agent, model, retrieval, and tool work in one trace using current OpenTelemetry GenAI conventions without exporting prompts or secrets by accident.

Turn an agent architecture into a data-flow model, identify attacker paths across tools and memory, and convert the highest risks into testable controls.

Build version-aware retrieval with valid-time ranges, immutable sources, time-filtered ranking, and citations that preserve what was known when.

A privacy-first OpenTelemetry Collector pipeline for retaining useful agent traces without exporting sensitive prompts by default.

Choose a multi-agent topology using a benchmark protocol that measures task success, coordination cost, latency, and unsafe actions.

Choose estimands, experimental units, paired comparisons, confidence intervals, power, and release thresholds that match stochastic agent evaluations.

Compare a candidate agent on representative production inputs while suppressing side effects, protecting user data, and avoiding misleading metrics.

Share one pinned base model across reviewed LoRA adapters while preserving lineage, tenant routing, cache isolation, memory limits, and measurable quality.

Configure vLLM's OpenAI-compatible server, model-specific parser, strict schemas, and a safe execution loop for local tool calling.

Authenticate, authorize, validate, rate-limit, and make high-risk agent mutations idempotent in Next.js 16.

Build a browser-worker boundary that treats webpages as hostile, isolates credentials, blocks unapproved network destinations, and requires approval for consequential actions.

Give agent tools short-lived credential handles while keeping passwords, tokens, keys, and connection strings out of prompts, memory, traces, and results.

Choose and harden an isolation boundary for generated code using explicit filesystem, network, process, syscall, resource, and output controls.

Give each coding agent an isolated branch and working directory, then integrate changes through explicit ownership, verification, and cleanup.

Test whether malicious search, document, email, and tool outputs can trigger unauthorized actions, memory poisoning, or data exfiltration.

Choose a RAG chunking strategy with a held-out benchmark that measures retrieval, answer evidence, citation quality, latency, and index cost.

Contain direct and indirect prompt injection with provenance, privilege separation, constrained tools, output validation, egress controls, and bound approvals.

Map multimodal data flows, collect only necessary regions and intervals, redact before model access, constrain retention, and verify deletion.

Build an outbound-request broker that stops agents from reaching internal services, cloud metadata, redirected targets, and DNS-rebinding destinations.

Control delegation depth, fan-out, duplicate work, cycles, authority propagation, and result synthesis in multi-agent systems.

Implement the MCP 2025-11-25 authorization boundary for an HTTP server without token passthrough, audience confusion, or excessive scopes.

Adopt Cache Components without leaking sessions, serving stale agent state, or invalidating more data than necessary.

Enforce tenant isolation across documents, chunks, vector retrieval, citations, caches, background jobs, and database roles.

Route agent calls by capability, quality, latency, cost, privacy, and availability using an evaluated policy with safe fallbacks.

Move a coding agent to a new framework through contract inventory, compatibility tests, dual runs, staged rollout, and a defined rollback.

A package-by-package migration plan for TypeScript 6 defaults, deprecations, type environments, project references, and the TypeScript 7 transition.

Turn raw agent events into reviewable memories with provenance, deduplication, contradiction handling, decay, and safe forgetting.

Instrument a voice agent with one clock model, turn-level milestones, WebRTC media statistics, and latency budgets that expose the real bottleneck.

A practical architecture guide to using MCP for tools and context, A2A for peer-agent work, and both without collapsing their trust boundaries.

Choose the right MCP primitive for actions, application-selected context, and user-invoked workflows, then design each contract safely.

Compare MCP's two standard transports and run a fair benchmark covering startup, latency, concurrency, streaming, recovery, and security.

Let an MCP server request model work through its client while enforcing tool capability checks, iteration budgets, approvals, and message balance.

Implement form- and URL-mode MCP elicitation with explicit navigation approval, schema validation, identity binding, and safe resume behavior.

Use dependency graphs, a tested worker pool, cancellation, deduplication, and budgets to reduce multi-agent latency without creating rate-limit and cost failures.

Calibrate a model grader with blinded human labels, bias probes, agreement diagnostics, and explicit rules for when automation must abstain.

Combine PostgreSQL lexical search with pgvector semantic retrieval using reproducible ranking, tenant-safe filters, and measured relevance.

Design approval previews, cryptographic action binding, expiry, authorization, execution, and audit trails for consequential agent actions.

A reproducible load-testing protocol for agent queues, models, tools, retries, side effects, quality, and recovery.

Build a voice-agent evaluation suite that measures task outcomes, transcript errors, turn behavior, interruptions, audio delivery, safety, and user control.

Define measurable reliability, latency, quality, safety, and cost objectives around user outcomes—not model uptime.