https://www.volansaquilae.com/guides/ Dated, sourced guides for teams putting AI agents in front of customers. en-gb Fri, 06 Nov 2026 09:00:00 +0700 https://www.volansaquilae.com/guides/production-ready-ai-agents/ https://www.volansaquilae.com/guides/production-ready-ai-agents/ Fri, 06 Nov 2026 09:00:00 +0700 Before an AI agent takes customer traffic, it should pass three gates: evals built from its own failures, a review of every tool and MCP connection against named threat lists, and a written go, go-with-conditions or no-go verdict that names an owner for every residual risk. Operational checks on monitoring, a kill switch and rollback, rate limits, human escalation and cost per conversation complete the review. https://www.volansaquilae.com/guides/ai-agent-deployment-checklist/ https://www.volansaquilae.com/guides/ai-agent-deployment-checklist/ Fri, 06 Nov 2026 09:00:00 +0700 An AI agent deployment checklist lists the checks an agent must pass before real users meet it, each with the evidence that proves it. This one has 41 checks in 11 sections, from scope and evals to tool permissions, prompt injection, the kill switch, rollback and cost. An unchecked item is a finding, graded as a blocker, a condition or a note. https://www.volansaquilae.com/guides/golden-dataset-from-production-failures/ https://www.volansaquilae.com/guides/golden-dataset-from-production-failures/ Fri, 06 Nov 2026 09:00:00 +0700 Build a golden dataset for LLM evaluation from failures your users have already hit. Mine production traces and support tickets, group the failures into a taxonomy, and write each one as a test case with an expected behaviour and a code grader. Freeze the cases as a versioned regression set and add a case whenever a new failure appears. https://www.volansaquilae.com/guides/llm-as-a-judge-calibration/ https://www.volansaquilae.com/guides/llm-as-a-judge-calibration/ Fri, 06 Nov 2026 09:00:00 +0700 Use an LLM as a judge only where a code check can't decide, and give each judge one pass/fail criterion for one failure class. Calibrate it before it gates anything. Hand-label 100 to 200 cases, write the prompt on some of them, and compare the judge's true positive and true negative rates on the held-out rest with a bar written down in advance. Report the judge's pass rate corrected for its measured error rates, with an interval, and measure again whenever the judge model, the prompt or the criterion changes. https://www.volansaquilae.com/guides/mcp-vs-api/ https://www.volansaquilae.com/guides/mcp-vs-api/ Fri, 06 Nov 2026 09:00:00 +0700 An API is an interface for software that developers write code against. MCP (Model Context Protocol) is an open protocol that lets AI applications such as ChatGPT and Claude find and call tools while they run. An MCP server usually sits on top of an existing API. It offers a few tasks described in words a model can act on, handles each user's sign-in, and speaks JSON-RPC. You need one when your customers want your product inside their AI tools. Your API is enough when developers, or your own agent, call it directly. https://www.volansaquilae.com/guides/mcp-server-security/ https://www.volansaquilae.com/guides/mcp-server-security/ Fri, 06 Nov 2026 09:00:00 +0700 MCP server security means your server accepts only tokens issued for it, never forwards them, asks for the narrowest scopes, binds state handles to the verified user and treats tool text as untrusted. The current specification, 2026-07-28, replaces sessions with state handles, adds issuer checks against mix-up attacks and deprecates Dynamic Client Registration in favour of Client ID Metadata Documents. The checklist below maps each risk to its spec section and a test you can run. https://www.volansaquilae.com/guides/mcp-oauth-cimd/ https://www.volansaquilae.com/guides/mcp-oauth-cimd/ Fri, 06 Nov 2026 09:00:00 +0700 MCP OAuth makes your MCP server an OAuth 2.1 resource server. It answers a tokenless call with a 401 that points to its protected resource metadata, and accepts only tokens issued for its own URL. On the 2026-07-28 revision, clients identify themselves with Client ID Metadata Documents, HTTPS URLs that serve their registration details, and Dynamic Client Registration is deprecated. Your authorization server fetches those documents safely, enforces S256 PKCE, returns iss and binds tokens to the resource the client names, and the conformance table in this guide tests each step. https://www.volansaquilae.com/guides/test-mcp-server/ https://www.volansaquilae.com/guides/test-mcp-server/ Fri, 06 Nov 2026 09:00:00 +0700 Test a remote MCP server in five layers: selected scenarios from the official conformance suite, contract and authorisation tests through the SDK's in-process client, tool-use evals with a real model on a schedule, and security tests against staging. Use MCP Inspector's web UI to explore a server and its CLI for smoke tests in CI. Pin the protocol era in every test, because the TypeScript SDK client and the Inspector both default to the 2025 handshake. https://www.volansaquilae.com/guides/mcp-error-codes/ https://www.volansaquilae.com/guides/mcp-error-codes/ Fri, 06 Nov 2026 09:00:00 +0700 MCP errors come in three kinds: JSON-RPC codes (-32700 to -32603), codes the SDKs use for local failures (-32000 when the connection closes, -32001 when a request times out) and HTTP statuses from the Streamable HTTP transport (400, 404, 405, 406, 415). Most trace back to four causes: a stdio server that exits or writes logs to stdout, a client and server on different protocol revisions, a session the server no longer holds, and an authorization server that offers no way to register a client. We reproduced the messages below on the official TypeScript SDK. https://www.volansaquilae.com/guides/mcp-stateless-migration/ https://www.volansaquilae.com/guides/mcp-stateless-migration/ Fri, 06 Nov 2026 09:00:00 +0700 MCP stateless migration moves a remote MCP server from the session-based 2025 revisions to revision 2026-07-28, which removes the initialize handshake, the Mcp-Session-Id header, the GET stream and server-sent requests. Each request now carries its own protocol version and client capabilities, so the work is reading that metadata on every request, implementing server/discover, returning Multi Round-Trip Requests, moving notifications to subscriptions/listen and adding resultType, ttlMs and cacheScope to results. The spec lets one endpoint serve 2025-11-25 and 2026-07-28 clients at the same time, so you can switch without cutting off older clients. https://www.volansaquilae.com/guides/ai-red-teaming-vs-penetration-testing/ https://www.volansaquilae.com/guides/ai-red-teaming-vs-penetration-testing/ Fri, 06 Nov 2026 09:00:00 +0700 A penetration test shows which technical vulnerabilities an attacker can exploit in the systems you scoped, and AI red teaming shows how a model or agent can be pushed into harmful or unintended behaviour. A production readiness review decides, against thresholds agreed in advance, whether one agent can take customer traffic. If the agent can change customer data or move money, we'd run the readiness review first and a targeted red team second, and book the penetration test your customers' security reviews ask for. https://www.volansaquilae.com/guides/owasp-agentic-top-10-tests/ https://www.volansaquilae.com/guides/owasp-agentic-top-10-tests/ Fri, 06 Nov 2026 09:00:00 +0700 The OWASP Top 10 for Agentic Applications, published on 9 December 2025, lists the ten highest-impact security risks in AI agents, from ASI01 Agent Goal Hijack to ASI10 Rogue Agents. Each can be tested before launch, with scripted attacks and eval cases graded on the agent's tool calls, or with configuration and log checks on credentials, pinned components, sandboxes and kill switches. Supply chain changes, memory poisoning, cascading failures, human over-trust and rogue behaviour also need monitoring after launch, because those risks keep changing in production. https://www.volansaquilae.com/guides/prompt-injection-test-cases/ https://www.volansaquilae.com/guides/prompt-injection-test-cases/ Fri, 06 Nov 2026 09:00:00 +0700 Prompt injection test cases check whether text an agent reads, from a user, an email, a document, a web page or another tool, can make it take an action nobody asked for. For an agent, grade each case on its tool calls, not its reply. A case fails if the planted instruction triggers a refund, an email, a share or a deletion. The 20 cases below cover direct, indirect and MCP routes, plus one control case, in JSONL you can run against your own agent. https://www.volansaquilae.com/guides/llm-regression-testing-model-updates/ https://www.volansaquilae.com/guides/llm-regression-testing-model-updates/ Fri, 06 Nov 2026 09:00:00 +0700 When a provider ships a new model or retires yours, keep production on a pinned model ID and re-run the full regression suite on both models, with several samples per case, in the same window. Compare the runs per failure class with a paired test and thresholds agreed in advance, check cost and latency, and switch through a staged rollout. Rollback means pinning the old ID again, which works only until its retirement date. https://www.volansaquilae.com/guides/llm-model-deprecation-dates/ https://www.volansaquilae.com/guides/llm-model-deprecation-dates/ Fri, 06 Nov 2026 09:00:00 +0700 Each provider lists its dates on its own lifecycle page, and Amazon Bedrock, Google Cloud and Microsoft Foundry keep separate schedules. Of the dates we last checked on 5 November 2026, the next big retirements are Gemini 3.6 Flash on Vertex AI and the seven o-series models on Azure OpenAI (19 November 2026), Claude Sonnet 4.5 on the Claude API, Google Cloud and Microsoft Foundry (30 November 2026), the GPT-5 and o3 snapshots and their aliases on OpenAI's API (11 December 2026) and the remaining GPT-4o and GPT-4.1 snapshots on OpenAI's API (14 January 2027). Plan to the earliest date on any platform you call. https://www.volansaquilae.com/guides/ai-vendor-security-questionnaire/ https://www.volansaquilae.com/guides/ai-vendor-security-questionnaire/ Fri, 06 Nov 2026 09:00:00 +0700 Enterprise buyers send AI vendors the Shared Assessments SIG, the Cloud Security Alliance's CAIQ or AI-CAIQ, or their own forms, and read the answers against ISO/IEC 42001, the NIST AI RMF, SOC 2 and the EU AI Act. For an agent, the questions fall into twelve areas, from training on customer data to tool permissions, prompt injection and evaluation evidence. This guide lists the questions in each area and gives an answer bank template with the evidence to attach. https://www.volansaquilae.com/guides/forward-deployed-engineer-cost/ https://www.volansaquilae.com/guides/forward-deployed-engineer-cost/ Fri, 06 Nov 2026 09:00:00 +0700 A US forward deployed engineer's median total compensation is USD 220,000 (Levels.fyi, 19 October 2026). That's about USD 258,000 a year with employer benefits and payroll taxes, or USD 302,000 in the first year with a 25% recruiter's fee, before management time and equipment (USD 328,572 fully loaded). The UK median is GBP 144,377, about GBP 161,000 a year with employer costs. Per month, a US agency contractor costs USD 22,400 to 38,400, a talent network about USD 25,000 and our white-label engineers USD 12,000.