Stanley Yang — Writing

A containment-first incident playbook for agent models, tools, prompts, memory, retries, and ambiguous side effects.

Turn an end-to-end response target into measurable queue, model, tool, orchestration, and streaming budgets.

Turn real agent tasks and failures into a versioned, reviewable evaluation dataset with defensible labels, provenance, and release coverage.

Find the prompt, model, tool, data, or code change that caused an agent regression despite stochastic outcomes and external drift.

A reproducible protocol for measuring a model router's quality, cost, latency, abstention, and resilience without cherry-picking.

Fuse pixels with roles, names, states, geometry, and postcondition checks while detecting disagreement between visual and semantic UI representations.

Compare graph and vector retrieval on local facts, multi-hop relationships, corpus-wide themes, latency, cost, freshness, and citation support.

Use planner, generator, and healer agents under explicit evidence, security, and review gates so test maintenance does not weaken assertions.

Store immutable agent events, build deterministic projections, preserve model and tool outcomes, and recover safely after failures.

Design deterministic before-and-after hooks that deny unsafe agent actions, require approval, verify patches, and create an auditable policy trail.

Choose AI backend placement with a reproducible latency, quality, capacity, cost, security, and failure benchmark.

Design long-running agent workflows that survive crashes, retry only safe work, bind approvals to exact actions, and reconcile ambiguous side effects.

Measure absent, concise human-written, generated, and noisy repository instructions with paired tasks, fixed agent settings, and no invented results.

Separate prefill and decoding only when independent scaling and latency isolation beat KV-transfer cost, duplicated weights, and operational complexity.

Record model, tool, time, randomness, configuration, and state dependencies so agent failures can be replayed without repeating real-world effects.

Turn design states into deterministic Playwright checks that combine semantic assertions, visual baselines, accessibility, and human-reviewed diffs.

Turn human approval into a real security control by previewing exact effects, binding approval to immutable action data, and preventing replay or substitution.

Use ratified WASI 0.3 async components with pinned runtime support, explicit capabilities, deadlines, backpressure, and safe cancellation.

Preserve user intent and actor identity across agent-to-service and agent-to-agent calls with narrow audiences, scopes, consent, and delegation chains.

A practical ledger for attributing model, tool, compute, retry, and shared costs to successful multi-agent outcomes.

Test agent tool schemas, transport behavior, authorization, retries, and compatibility with valid, invalid, and adversarial requests.

Design a context pipeline that selects trusted instructions, relevant evidence, tool definitions, and recent state within an explicit token budget.

Choose between APIs, structured browser automation, accessibility trees, and visual computer use using reliability, coverage, security, and cost.

Select open weights with a reproducible, product-specific suite covering capability, tool use, safety, latency, memory, and operational fit.

Replace broad ambient credentials with narrow, expiring, attenuable grants bound to one agent task, tenant, operation, and resource set.

A correctness-first guide to caching model prefixes, retrieval results, tool calls, and final AI outputs safely.

Connect a coordinator to an independently operated report agent using an Agent Card, SendMessage, task state, typed artifacts, and strict authorization.

Stream agent UI and typed run events with Suspense, Route Handlers, cancellation, backpressure, and accessible status updates.

Design a durable MCP tool call with polling, deferred results, input-required handling, cancellation, and tenant-safe retention.

Capture actor chains, policy decisions, approvals, tool effects, and evidence so investigators can reconstruct an agent incident without logging hidden reasoning or secrets.

Turn the OWASP Top 10 for Agentic Applications into safe, repeatable abuse cases with simulated tools, hard invariants, evidence, and CI gates.

Connect to a local MCP server, verify capability discovery, call a typed tool, and handle remote transports without confusing discovery with authorization.

Design an MCP tool that serves an interactive UI resource, receives structured results, and stays safe across hosts with different extension support.

Design a review agent that reports reproducible defects with precise evidence instead of producing noisy summaries and style opinions.

Define typed WIT contracts, explicit capabilities, resource limits, provenance, and conformance tests for portable agent tools.

Create resettable visual web tasks with programmatic outcome validators, layout perturbations, safety checks, and reproducible multimodal traces.

Reserve one GPU for generation, run embeddings and retrieval on CPU, store cited chunks in Postgres with pgvector, and evaluate retrieval before answers.

Create resettable, observable email, calendar, browser, and database environments that test agent effects without touching real systems.

Build and test a real MCP stdio server, then make the transport, validation, error, and authorization decisions required for a remote deployment.

Design a reconnectable browser voice agent with explicit signaling, ephemeral authorization, audio lifecycle, interruption, and observable session state.

Design a restartable RAG ingestion pipeline that preserves provenance, detects changes, avoids duplicate work, and removes stale chunks from retrieval.

Implement an agent loop with typed plans, evidence-based review, bounded revision, explicit stop states, and safe tool execution.

Route remote MCP traffic by authenticated tenant while isolating tokens, sessions, discovery, quotas, results, and audit records.

Estimate weights, KV cache, runtime overhead, context concurrency, and replica demand—then replace assumptions with measured service curves.

Wrap any coding model in a reproducible harness that controls workspaces, commands, credentials, budgets, evidence, and verification.

Represent evidence at span level, require citations for verifiable claims, and abstain when retrieval cannot support a grounded answer.

Preserve page regions, reading order, tables, OCR alternatives, and source versions so every document answer can resolve to visible evidence.

Enforce token, tool-call, latency, concurrency, and monetary ceilings across nested agent work without relying on prompt instructions.

A reproducible protocol for measuring JSON validity, schema compliance, semantic accuracy, latency, and failure recovery across open models.

Measure speculative decoding against a target-only baseline across load, prompt types, acceptance rates, memory, latency, throughput, and output equivalence.