What's new · newest first
Changelog
The changes that affect how you learn here — new features, improvements, and fixes — in plain language. We update the platform continuously as the field moves; this records the changes worth telling you about, not every deploy.
August 2026
latest-
Improved
A worked example we taught was itself the mistake it was warning about
The observability guide and its module taught you to rank failure modes by how loud they are, and used a timestamp bug as the quiet-but-wrong case: truncating a millisecond value to ten digits, supposedly giving a date wrong by decades that still looks fine on a dashboard. It does not. Cutting the low-order digits off a base-10 number is division by a power of ten, so those ten digits are exactly the right answer. The lesson was correct and its example was not — which is the failure it describes. Both now use a real one: a queue roll-up that counted already-decided rows as backlog, so settled work appeared to age one day per day.
-
Improved
Six guides now show what actually broke here, not just how the idea works
The agent-security, prompt-injection, observability, cost, agent-cost and evaluation guides gained sections drawn from this codebase: a stored self-XSS round trip, a capability latch set from the whole tool array before dispatch, a thirteen-day pin on a wrong error code, a dry run that caught 680 characters where 261 were intended. Each names the file it came from. The prompt-injection guide also states both failure modes of our own guardrails, including what stops being checked when the screening service degrades.
-
New
Two issues of the Brief went live: MCP going stateless, and a Cloudflare week worth reading in one sitting
Issue #005 covers the 2026-07-28 MCP revision that removed the initialize handshake — including the four gaps we found in our own client when we audited it against the published spec — plus the EU AI Act dates that arrived on 2 August and OpenAI's price cut. Issue #006 covers Cloudflare's Agents Week, AWS giving AgentCore agents fourteen-day sessions, and the Sonnet 5 introductory price ending on 31 August. #005 was written for the week of 3 August and published late; both went live together, and the RSS feed carries them at their real publication date rather than back-dated.
-
Improved
A vendor rate we published was six times too high, and the parity claim underneath it was missing a condition
The cost pages quoted the wrong floor for Workers AI input tokens — the figure was a real published rate, just not the cheapest model on the list, with several below it. Corrected wherever it appeared, and the agent cost guide now names the model each end of the range belongs to. The same pass found a broader problem: we said per-token cost is identical on the Anthropic API and Bedrock, unqualified, in twelve places, and it is only true on Bedrock's global endpoints. Regional endpoints and geographic inference profiles carry a premium — and so does the first-party API once you pin inference to a named geography, which means data residency costs the same either way. Every instance now says so.
-
Improved
AWS raised the ceiling on how long an agent session can run, so the curriculum stopped calling eight hours the limit
Bedrock AgentCore's managed runtime caps a session at eight hours, and we taught that as the ceiling. On 6 August AWS made runtime instances generally available: attach an agent to your own EC2 capacity and a session can hold for up to fourteen days, billed as EC2 plus a management charge. Eight hours is still right for the default serverless path, so this was a missing qualifier rather than a wrong fact — but an unqualified ceiling is the kind of thing an architect plans around. Both the Track D lesson and the public comparison of where to run Claude agents now name both paths.
-
Improved
Sixteen new worked examples from this platform's own build
Twelve lessons across Tracks B, C, D and E gained real examples from building this site — a retrieval floor that silently returned nothing, a report-only security policy that stayed quiet right up until it killed the page, a gateway failure mode that looks like success. Most name the file, commit or post-mortem they came from, so you can go and read the thing that actually happened. Thirteen further candidates were set aside, mostly for restating something the module already taught.
-
Improved
The human-in-the-loop pattern was two permission modes short, and wrong about a third
Claude Code documents six permission modes; our pattern page described four. The two missing ones matter more than a count suggests —
autoputs a model classifier in the seat the human used to occupy, anddontAskturns every prompt into a denial. The page also saidplanforbids execution, when read-only tools run normally under it. Corrected against the vendor documentation, along with a related point the page had right and our own notes had backwards: a permission check incanUseToolis bypassed entirely for any tool auto-approved earlier. -
Improved
Two vendor facts moved under our pages, and the pages moved with them
The MCP specification's versioning page flipped the revision it calls "current" from
2025-11-25to2026-07-28— we had published, correctly, on 28 July that it had not; the pointer flipped on or before 3 August. Seven places across the MCP guides, the integration workflow and two lessons said so in the present tense, and all seven are now re-mapped:2026-07-28is current,2025-11-25is Final rather than deprecated, and a server speaking only the older revision still works because version is negotiated per request. Separately, AWS closed nine SageMaker AI features to new customers on 30 July — including Ground Truth, Model Monitor, Clarify and A2I, which Track D taught without a note. Those lessons now carry the availability status alongside the mechanics, because the exam still tests them even though a new account cannot adopt them. Dated corrections stay visible on the Brief issues rather than being edited away. -
New
A guide to the Forward Deployed Architect role — and the certifications sold for it
Forward Deployed Architect is a real job title: Palantir, Google Cloud, NVIDIA, Adobe and DevRev all post under it, and it pays well. A certification market has grown up around the role very quickly. The new guide sets the two side by side — what the job asks of a senior engineer, what the published figures actually say, and why the certifications sold for it come from third parties rather than from any of the companies doing the hiring. It sits alongside the existing guide to the forward deployed engineer role, and the two now link to each other.
July 2026
-
Improved
Fonts are now served from our own infrastructure — and the body typeface finally loads
Every page used to fetch its typefaces from Google's font CDN, which meant a third-party request on every visit before any text could render. All three families now come from our own object storage, same-origin, and the Content Security Policy no longer permits Google's font domains at all. Fixing that surfaced a longer-standing bug: the stylesheet asked for a family named "Hanken Grotesque", and the real family is "Hanken Grotesk" — a font request for a name that does not exist is dropped silently rather than refused, so the body typeface had never once loaded and every page had been falling back to whatever sans-serif your operating system supplies. Prose across the site now renders in the face it was designed in.
-
Improved
Every MCP page and lesson updated for the 2026-07-28 protocol revision
The Model Context Protocol published its largest revision since launch: the
initializehandshake is gone, protocol-level sessions and theMcp-Session-Idheader are removed, three request headers become mandatory on Streamable HTTP, and Tasks moved out of the core spec into an extension. Both MCP modules and all five MCP pages — what MCP is, how to build a server, MCP vs an API, the integration workflow and agentic RAG — now teach both eras rather than quietly assuming the old one. The detail most write-ups were getting wrong at the time: the revision was published while the spec's own versioning page still named2025-11-25the current protocol version. That flipped on or before 3 Aug 2026 —2026-07-28is now named current,2025-11-25is Final, and a server speaking only the older revision is still conformant because version is negotiated per request. Our pages carry both facts, plus the compatibility matrix that explains why your existing server keeps working — and the asymmetry that a legacy client does not. -
New
Workflow: running several AI agents on one repository without losing work
A new page in the Engineering Workflow Library: parallel agents, one working tree. When several agents write to the same checkout at once, an overwritten edit produces no error, no failing test and no conflict marker — the successful write destroys the evidence that anything was lost. The page carries four dated collisions from building this platform, including the one that started it: seven agents with deliberately non-overlapping file lists, and one of them still reverted another's work. The file lists were the wrong idea. What collides is the kind of write each agent performs, not where it happens, and a bulk find-and-replace pass legitimately owns every file it matches no matter who else is editing there. Seven steps, a failure-mode table, and the honest comparison against giving each agent its own checkout instead — including what that costs.
-
Improved
Sources now sit next to the claims they support
Across 24 library and guide pages, citations that lived only in the provenance block at the foot of the page are now linked inline, on the sentence making the claim — so you can check a figure where you read it instead of scrolling to a list and matching it up yourself. The provenance blocks stay. This came out of a check on our own tooling: the gate meant to require citations was counting the ones in that end block, which meant a page could satisfy it while having nothing beside any actual claim. One page is deliberately left alone rather than patched — its two sources genuinely do not support anything its text says, and attaching them to something adjacent would have been worse than leaving the gap visible. Every page also now carries links to the pattern, workflow and pricing libraries in its footer, which previously existed only in a menu built by JavaScript and so were invisible to anything that does not run it.
-
Improved
Two pages printed a command that does not exist
The deterministic gate pattern and the spec-first workflow both showed a hook configuration calling a script file that has never existed in this repo. Anyone copying it would have got a silent no-op — the exact failure both pages are about. Both now show the real mechanism, including the part that matters: it injects a reminder rather than blocking the edit, and blocking needs a different exit code. The spec-first page also gained a section listing the real files this practice runs on, and a spec-template section list that was one item short.
-
New
Pattern: an empty state must not render as a value
A new pattern page, fail empty, on a failure that arrives disguised as two unrelated bugs. A banner whose consumer overwrites placeholder text but never clears it shows one visitor a result belonging to nobody; a relevance floor written
score < limitkeeps a hit whose score came back malformed, because every comparison with a non-number is false — so the guard evaporates in exactly the conditions it exists for. Both are the same missing decision about what absence should look like, and in both the obvious fix is the wrong one: hiding the placeholder loses to a stylesheet rule, and sharing one constant across two branches that measure different things is coupling wearing DRY's clothes. It covers the write-time form that makes the safe side explicit, the counter-example already correct in this build, and when not to reach for it. Both instances are dated defects from this platform's own build, including the fix that was wrong first. Free to read — no sign-up. -
Improved
Ask the Guides now names every source it read
When the guide assistant answers from the library, it lists the guides it was given — not just the closest two. Previously it retrieved up to three passages and cited at most two, so a guide could shape an answer without being named. Answers that fall back to the deterministic reply still cite only what that reply actually quotes, because over-citing is the same problem in reverse.
-
Improved
Practice questions you cannot answer by pattern-matching
A sweep of the Bedrock and data-pipeline question banks removed answer cues that let you score without reading the scenario — options where the right answer was simply the longest, or the one naming the most managed services, or the only one echoing a phrase from the question. On the affected set, guessing by “most managed services” dropped from 77% to 3% against 25% chance. One answer key turned out to be wrong and is corrected.
-
New
Pattern: when a rule is checkable, write the check
A new pattern page, the deterministic gate, on what to do the second time a rule you have already written down gets skipped. Keep the instruction — but move the part that is mechanically decidable out of it and into a check that runs as code and fails closed on the one path to production. It covers the three layers of the pattern, the friction it costs you, the rules you must not gate this way, and the exit-code trap that leaves a gate looking green while it checks nothing. Free to read — no sign-up.
-
New
Workflow: fixture the first check before you trust it
A new workflow, validating the validator, on the check nobody checks. Write the first validator for a file everything downstream trusts and its verdict inherits that trust — so a green is indistinguishable from a check that never ran, and a red arrives looking exactly like a real defect in the file rather than a bug in the checker. The practice: read the raw bytes before you write the check, write the awkward known-good fixture first, give the check a way to run against something other than production, and calibrate a threshold instead of asserting one. Three dated failures from this platform's own build, including one where a broken checker's report was filed, prioritised, and killed at the last gate before authoring. Free to read — no sign-up.
-
Fixed
Corrected how this changelog describes access and plans
Several entries on this page described access, plans and automation in the present tense and had not been revisited since they were written. A date heading bounds when a change shipped, not how long its wording stays true. Free access to the foundational track was offered here to "every signed-in learner" — accurate when it was written, but sign-up has been waitlist-only since 11 July 2026, so there is no public route to an account today beyond the waitlist; what anyone can open with no account is the sample lesson. The paid plan was called a "Pro subscription"; it is the Professional Membership, and there is exactly one. And the price tracker's rates were said to refresh automatically — a routine change does publish on its own, but an outsized swing, or a model we have not tracked before, is held for review first. Those entries now say what was true then and what is true now, and this entry records the correction rather than leaving the edits silent.
-
Fixed
Corrected how we describe AI Gateway guardrail modes
Eighteen pages across the guides, pattern and workflow libraries said, or implied, that our Cloudflare AI Gateway guardrails run in flag mode. That was true of one hazard category and wrong about the rest: prompt injection and jailbreaks run in flag mode deliberately — block mode flags our own security curriculum as an attack — while the other thirteen categories block on both the prompt and the response side (the coach's streamed replies are covered prompt-side only). The distinction matters, because the two modes fail in opposite directions: block fails closed if the evaluator is unavailable, flag proceeds without evaluating. Pages now say which mode applies where, and the founder log carries a dated correction rather than a quietly edited entry.
-
Improved
Deeper question bank for Claude Code configuration
The Claude Code configuration module in Track C now carries twenty-six practice questions instead of thirteen, covering the third domain of the Anthropic Claude certification (CCAR-F) more thoroughly. Every item cites its source, as all of ours do — the point is a bank deep enough that repeat attempts test recall rather than memory of the last attempt.
-
Improved
Model references updated for Claude Opus 5
Anthropic released Claude Opus 5 on 24 July 2026. The calculator on what a Claude agent costs now models it, and the model-selection, routing and cost guides name it where it changes the answer. Opus 4.8 stays wherever it is still the right referent — it has not been retired, and worked examples built on it keep their original figures rather than being silently relabelled.
-
New
Guide: adding a model is a data-flow change
A new guide at /llm-subprocessors on a question that is easy to miss when you swap a model in a routing config: who processes your users' data now. Open-weight models decouple the publisher from the provider actually serving the weights, so the answer depends on the serving path rather than the model card — and the guide shows how to find it, which contract clause settles it, and why the check belongs in the same change as the routing edit. Free to read — no sign-up.
-
Improved
A context-aware safety check on the AI coach
The coach now checks your messages for genuine attempts to override its instructions before it answers, using a classifier that sees only your own recent messages — never the lesson content. That last detail is the point: a plain content filter cannot tell an attacker probing prompt injection from a member studying it, and this platform teaches the subject, so questions about how these attacks work are not what it looks for. It is one layer among several, not a guarantee — and when the check itself is unavailable, the coach answers anyway rather than blocking you. We also expanded the write-ups behind it on defense in depth, red-teaming LLM systems, AI-assisted review, and AI guardrails.
-
New
Role guides: AI engineer, AI developer, AI solutions architect
Three new guides — /ai-engineer, /ai-developer, and /ai-solutions-architect — covering what each role actually does day to day, how the three differ from one another and from the AI architect, the skills and certifications that carry weight, and how an experienced software engineer moves into each one. Sourced, and free to read — no sign-up.
-
New
Live LLM API pricing, in one place
A new price tracker at /llm-pricing shows the current per-token input and output rates for the major model providers — Anthropic, OpenAI, Google, xAI, DeepSeek, and leading open-weight models like Kimi, GLM, and Qwen — side by side. It comes with a cost calculator and a free, no-auth JSON API you can pull into your own tooling; every rate cites its source, and a routine price change publishes on its own, while an outsized swing, or a model we have not tracked before, is held for review first.
-
New
The Engineering Workflow Library
In-depth, sourced playbooks at /workflows for building with AI as a senior engineer — including Claude Code, agent-SDK delivery, MCP integration, context engineering, AI-assisted code review, eval-first development, the AI quality stack, spec-first development, red-teaming LLM systems, and a security self-audit. Free to read, no sign-up.
-
New
The Architecture Pattern Library
Production-grade AI architecture patterns at /patterns — including the bounded agentic loop, grounded RAG, model routing, eval-harness gates, human-in-the-loop approval, agent memory, defense-in-depth, and the capability latch — each written up with the design forces, trade-offs, and real as-built evidence behind it.
- New
June 2026
-
New
A deeper AWS Generative AI Developer track
Our preparation track for the AWS Certified Generative AI Developer – Professional exam grew substantially. The heavier topics — foundation-model deployment, AI safety & governance, and cost & performance — are now split into focused modules, each with its own objectives and fresh scenario practice: when to reach for asynchronous inference on large GPU jobs, enforcing guardrails at the permissions layer, tracing which safety policy blocked a request. Every question is sourced to the official AWS docs.
-
Improved
Diagrams that make the hard concepts click
Several of the more spatial guides — embeddings, multi-agent orchestration, context engineering, RAG vs fine-tuning, and agentic design patterns — now open with a clean, on-brand diagram that shows the idea at a glance before you read it. Same restrained style as the rest of the site, no stock-art.
-
Improved
A cleaner experience on your phone
We tuned the site for small screens: the homepage menu no longer crowds the edge, the wordmark and page headers sit on one tidy line, and wide code samples and comparison tables now show a soft fade at the edge so it's clear you can scroll them sideways for more.
-
New
Privacy-first analytics, only if you say yes
We added optional analytics (Microsoft Clarity) to understand how the site is used and improve it. On your first visit you'll see a short banner: nothing loads and no analytics cookie is set unless you accept, declining is just as easy, and you can change your mind anytime from the cookie policy.
-
Improved
Stay signed in — no surprise re-logins
Your session now stays valid while you work, and a brief connection hiccup no longer kicks you out of a lesson or drops an answer you just submitted. If you do need to sign in again, you land back on the exact page you were on.
-
Improved
The site knows when you're signed in
Public pages now recognise a signed-in account and send you straight to your dashboard instead of asking you to log in again.
-
Improved
A smoother start on the foundational track
Signed-in accounts with no paid plan began straight away with the foundational Track 0 modules open to them, while the full placement stayed visible as a goal, and locked modules showed a clear "upgrade to unlock" prompt instead of an error. That still holds for an account that has it — but sign-up has been waitlist-only since 11 July 2026, so it is not something a visitor can reach today. What is open with no account is the sample lesson.
-
New
Your profile page
A new profile page brings together your account summary, learning stats, goals, and email preferences in one place.
-
New
Membership and in-app upgrade
Membership unlocks the full curriculum and the AI coach, and the upgrade flow moved in-app. It shipped that June as a "Pro subscription" and is now the Professional Membership — there is exactly one. The foundational track stays open to a signed-in account with no paid membership, but under waitlist-only access that is an entitlement on an existing account, not something a visitor can sign up for.
-
New
Portfolio submissions
Members submit a real artifact — a repository, a pull request, a demo, or a write-up — as proof of the applied work behind a module, and it comes back reviewed. Verifiable work, not self-attestation.
-
New
Program progress on your dashboard
A progress card shows how many modules you've mastered across every track — framed as competence, not hours logged.
-
New
Dozens of new free guides
The free guides library grew by more than 40 pages — agentic tooling and frameworks, MCP, security and governance, vector databases, certifications, and career paths into AI engineering. Each is a sourced, no-sign-up reference, and every fact was checked against a primary source.
-
New
Free build-in-public architecture notes
A new set of free, sourced architecture notes joins the guides library — real engineering decisions from building this platform.
-
Improved
Routine security and maintenance updates
Ongoing behind-the-scenes updates to keep the platform secure and dependable.
May 2026
-
New
A new track: AI regulation & compliance
The curriculum gained a track on AI regulatory and compliance architecture — risk-tiering, the EU AI Act, US rules, and privacy-by-design — aimed at engineers who need to build to the rules.
-
New
Free guides library and public site
A free, sourced reference library for senior engineers moving into AI went live, alongside the published curriculum and a sample lesson — no sign-up required.
-
Improved
Accessibility and polish
Across-the-board refinements to colour contrast, motion, and navigation (breadcrumbs and a back-to-top control) for a clearer, more accessible experience.
-
New
Platform launch
The learning platform went live: lessons with quizzes, spaced-repetition review that brings concepts back just before you'd forget them, mastery tracking per objective, and an AI coach that helps you reason through a problem rather than handing you the answer — with sign-in and a personal dashboard.
Want these updates as you learn?
The platform keeps moving as the AI stack does. Browse the curriculum, or start with a free sample lesson.