Changelog
Every change to this site, including what was proposed and rejected. Parts of this site are maintained by AI agents; this page is their audit trail.
v0.5.24
- Published field notes for the week of October 5, 2026: Sierra and Meta's proposed Personal Agent Protocol, Decagon's open-sourced PACT and its Gartner Cool Vendor mention, Claude Haiku 5.5, OpenAI's wider Ultrafast rollout for GPT-6.1 Sol, Intercom's Macro Admin Mode, and DeepL's new EU-sovereign translation tier — seven items, each checked against its own source page. No reference entries were due for freshness re-verification this run (the oldest
verifieddate across both guides is 41 days old; the threshold is 60).
Verified against: https://sierra.ai/blog/introducing-personal-agent-protocol · https://decagon.ai/blog/introducing-the-personal-agent-consent-trust-protocol-pact · https://decagon.ai/blog/decagon-named-a-gartner-cool-vendor · https://www.anthropic.com/claude-haiku-5-5 · https://developers.openai.com/api/docs/changelog · https://www.intercom.com/changes/en/159696-macro-admin-mode · https://www.deepl.com/en/blog/building-for-europes-sovereignty-needs
v0.5.23
- Published "The Weekly Reading" in English and Spanish: the recurring hour where two people read what the AI support agent actually said to customers, what the reading catches that the queue numbers do not, and why the loop must never come off.
v0.5.22
- Field notes, week of September 28, 2026: Anthropic committed $100 million to launch Claude Frontier Academy, training 10,000 "Frontier Deployed Engineers" by end of 2027; OpenAI added computer use to its Agents API and released GPT-6.1 Sol plus an ultrafast service tier for GPT-6 Astra; Intercom shipped duplicate-profile merging and larger, more accessible tooltips; Sierra and Decagon both became launch partners in OpenAI's new B2B marketplace; Lokalise published an overview of its enterprise integrations, citing a 93% turnaround-time cut at Navan; and Crowdin rounded up its September updates, including a Copilot "Ultracode mode" and an AI pipeline that routes only ambiguous strings to human review.
- Reference-refresh quota: no entries in data/reference/ were due this run — the oldest
verifieddate across both guides is roughly 34 days, under the 60-day threshold. The quota resumes once entries age past it.
Verified against: https://www.anthropic.com/news/claude-frontier-academy · https://developers.openai.com/api/docs/changelog · https://www.intercom.com/changes/en/158536-merge-duplicate-users-into-one-profile · https://sierra.ai/blog/sierra-joins-the-openai-marketplace · https://decagon.ai/blog/decagon-and-open-ai · https://lokalise.com/blog/enterprise-localization-integrations/ · https://crowdin.com/blog/whats-new-at-crowdin-september-2026
v0.5.21
- Weekly audit: trimmed two Spanish meta descriptions that exceeded 160 characters — "Measure the Queue, Not the Bot" (162 chars) and "It Says It Fixed It" (162 chars) — down to 150 and 157 respectively, with no change in meaning. All 97 external links across en/, es/ and data/ were checked; everything resolved (GitHub, OpenAI and LinkedIn returned non-200 responses to a scripted fetch — confirmed by hand as bot-blocking on well-known live sites, not dead links). No reference entries were due for freshness re-verification this run (oldest
verifieddate is 31 days; threshold is 60). Parity spot-check of three recently-changed pairs (field notes week of Sept 21, the draft "Multilingual Support Used to Be a Staffing Problem", and vibe-coding) found no drift. Publication hygiene swept clean — no leftover review markers, no stale draft flips, no placeholder mirrors. SEO basics otherwise pass (titles and feeds valid; the two over-length descriptions above were the only issue). No structural findings this run.
Verified against: https://example.com · https://www.anthropic.com/news/claude-opus-5-5 · https://developers.openai.com/api/docs/changelog · https://sierra.ai/blog/your-agent-laid-bare-and-why-it-matters · https://decagon.ai/blog/growing-in-benelux · https://www.intercom.com/changes/en/158136-build-the-metric-you-need-across-your-full-history · https://www.deepl.com/en/blog/deepl-available-in-microsoft-copilot-chatgpt-claude-mcp · https://github.com/amitkvint · https://github.com/amitkvint/atlasti-guided-coding · https://openai.com/ · https://www.linkedin.com/in/amitkvint/ · https://www.linkedin.com/in/ofwald/
v0.5.20
- Field notes for the week of September 21, 2026: Anthropic's Claude Opus 5.5, OpenAI's GPT-6 Sol/Luna release, Sierra's transparent-agents post, Decagon's Benelux expansion, two Intercom reporting/telephony changes, and DeepL's new MCP integration with Copilot, ChatGPT and Claude.
- Reference refresh: no ai-support-agents.json or ai-localization.json entries were due for the 60-day freshness quota this run (all under 60 days since the last check). While sweeping sources, updated the DeepL entry in ai-localization.json to note its new Model Context Protocol integration, re-verified against the DeepL blog post, and refreshed the guide's updated date.
Verified against: https://www.anthropic.com/news/claude-opus-5-5 · https://developers.openai.com/api/docs/changelog · https://sierra.ai/blog/your-agent-laid-bare-and-why-it-matters · https://decagon.ai/blog/growing-in-benelux · https://www.intercom.com/changes · https://www.intercom.com/changes/en/158136-build-the-metric-you-need-across-your-full-history · https://www.intercom.com/changes/en/157743-see-why-an-outbound-call-didn-t-connect · https://www.deepl.com/en/blog/deepl-available-in-microsoft-copilot-chatgpt-claude-mcp
v0.5.19
- Weekly audit: fixed a dead reference link. The Freshworks Freddy AI entry's url and source both pointed to freshworks.com/freddy-ai/ and /whats-new/, which now 404. Replaced both with the current freshworks.com/platform/freddy-ai/ product page, re-verified the summary against it, and refreshed the guide's updated date. All other external links across en/, es/ and data/ (91 checked) resolved; no reference entries were due for freshness re-verification this run (all under 60 days). Parity spot-check of three recently-changed essay pairs (vibe-coding, it-says-it-fixed-it, what-should-never-be-automated) found no drift. SEO basics (titles, descriptions, sitemap, feeds) all pass. No structural findings this run.
Verified against: https://www.freshworks.com/freddy-ai/ · https://www.freshworks.com/whats-new/ · https://www.freshworks.com/platform/freddy-ai/
v0.5.18
- Published "Vibe Coding a Live WordPress Site With No Fancy Tools" in English and Spanish.
- Accuracy pass on the same essay, the same day: the paid tools already in use are acknowledged, the claim about an agent editing the parent theme is stated as a likelihood, a paragraph now says who the wp-admin plugins do suit, the WP-CLI line notes it needs SSH, the script count is updated to eleven, and three lines about the working sessions that the project history could not confirm were rewritten to describe only what the files show.
Verified against: https://www.hollywoodreporter.com/lists/best-international-film-schools/escac/ · https://wordpress.tv/2012/12/13/mis-10-pasos-fundamentales-para-crear-temas-hijos/
v0.5.17
- Removed a stale duplicate of "What Should Never Be Automated". A CMS save made after the September 21 rename recreated the essay under its old drafting-date filename, with the pre-review text, and the build stopped on the duplicate address. The published version is unchanged.
v0.5.16
- Dated "What Should Never Be Automated" to its publication day, September 21, instead of its drafting date. Both language files and the linkedin/ sidecar were renamed to the new date prefix; the URL is unchanged in both languages.
v0.5.15
- Published "What Should Never Be Automated" with the approved edits restored. The CMS publish had overwritten them with an older copy: the grammar pass, the categories paragraph, the passage on HR and management looking at headcount, and the Spanish mirror of all of it. The build refused that publish because the older Spanish copy still had review markers, so it never went live.
v0.5.14
- Field notes, week of September 14: Anthropic's on-demand compaction for the Messages API; Sierra's AIUC-1 certification for its chat and voice agents; Decagon's GEPA-GAN adversarial method for testing agents and its new Generative Analytics feature; DeepL Voice's move to general availability with cross-language tone preservation; Lokalise's Vantage workspace for marketing content; a Crowdin case study on Sinch's 65% AI-automated translation; and WPML 5.0's unified admin and pay-per-page billing. Published in both languages.
- Checked all 18 reference entries in data/reference/ against the 60-day freshness quota: all fall within the window (oldest verified September 1), so none needed re-verification this run.
Verified against: https://platform.claude.com/docs/en/release-notes/api · https://platform.claude.com/docs/en/build-with-claude/compaction · https://sierra.ai/blog/sierra-achieves-aiuc-1-certification · https://decagon.ai/blog/gepa-gan-teaching-ai-to-sound-human · https://decagon.ai/blog/generative-analytics · https://www.deepl.com/en/blog/deepl-voice-preserves-your-voice-across-languages · https://lokalise.com/blog/introducing-vantage/ · https://crowdin.com/blog/sinch-localization-with-crowdin-and-undertow · https://wpml.org/announcements/2026/09/wpml-5-0/
v0.5.9
- Field notes, week of September 7: Sierra's multimodal agents that switch between voice, text and embedded visuals; Decagon's governance model for its self-improving Autopilot plus the DuetBench-2 results behind it; Intercom's pre-built Fin connector templates and CSV export for saved Inbox Views; Anthropic's
autopermission-policy mode for Managed Agents; Lokalise's Proof of Value feature for scoring AI translation against a team's own approved content; and a Crowdin case study on EMCD's move from 5 languages to 25 locales. Published in both languages. - Checked all 18 reference entries in data/reference/ against their
verifieddates: the oldest is September 1, 2026, so none has crossed the 60-day freshness threshold and none needed re-verification this run.
Verified against: https://www.intercom.com/changes/en · https://platform.claude.com/docs/en/release-notes/overview · https://support.zendesk.com/hc/en-us/sections/4405298877338 · https://www.freshworks.com/whats-new/ · https://help.salesforce.com/ · https://www.hubspot.com/new · https://sierra.ai/blog · https://decagon.ai/blog · https://www.deepl.com/en/blog · https://phrase.com/news/ · https://lokalise.com/blog/ · https://www.smartling.com/blog/ · https://crowdin.com/blog · https://lilt.com/blog · https://wpml.org/category/announcements/ · https://developers.openai.com/api/docs/changelog · https://www.intercom.com/changes/en/156987-connect-fin-to-your-tools-without-building-a-connector-from-scratch · https://www.intercom.com/changes/en/156366-export-a-saved-view-exactly-as-you-see-it · https://sierra.ai/blog/multimodal-agents · https://decagon.ai/blog/autopilot-in-production · https://decagon.ai/blog/duetbench-2 · https://crowdin.com/blog/emcd-localization-with-crowdin · https://platform.claude.com/docs/en/managed-agents/permission-policies · https://lokalise.com/blog/prove-ai-translation-quality-on-your-own-content/
v0.5.13
- Voice rules, round 3: six editing patterns observed in Amit's edit of a LinkedIn post, so future drafts start closer to how he writes. Examples come from real support work, vendor releases are described without praise, a closing question is left unanswered, and a light :) can appear early.
v0.5.12
- Retitled the September 10 field notes to "week of September 3, 2026". It shared the title "week of September 7" with the September 14 digest; every other digest is titled from the seven days before its date, and all its items fall between September 3 and 10.
- Corrected the v0.5.9 changelog entry, which said the field-notes run checked 8 reference entries. data/reference/ holds 18 (8 in the AI-localization guide, 10 in the AI support agents guide). The outcome is unchanged: the oldest verified date is September 1, 2026, so none was stale.
v0.5.11
- Restored two approved lines in "It Says It Fixed It. It Didn't." that the CMS publish overwrote with an older copy: the hardware passage (now the confirmed line about building PCs as a teenager, replacing a review marker) and the reference to the Tuesday essay, which again says "last week" and links to it. The build had refused the publish because of the marker, so the older copy never went live.
- Dated the essay to its publication day, September 14, instead of its drafting date. Both language files and the linkedin/ sidecar were renamed to the new date prefix; the URL is unchanged in both languages.
v0.5.10
- Dated "The Queue Looked Fine on Monday" to its publication day. It had kept its drafting date of September 2, which sorted it below essays from September 3, 5 and 11. Both language files and the linkedin/ sidecar were renamed to the new date prefix; the URL is unchanged in both languages.
- Restored the closing period of the English description, lost when it was shortened to fit the CMS's 160-character limit.
- Linked Otto Wald's name in the Spanish mirror, matching the English essay.
v0.5.8
- Corrected the setup section of the ATLAS.ti essay. It said the data does not leave your machine. Only the connection is local: the passages the assistant reads are sent to the model provider and processed under its terms. The essay now says so, and asks readers to check their data agreement and ethics approval first, matching the Data section of the services page. The unpublished LinkedIn version dropped the same claim.
v0.5.7
- Dated "The Agent Was Wrong All Tuesday" to its publication day. It had kept its drafting date of September 3, which sorted it below the ATLAS.ti essay of September 5 and made it look as though it had never gone live. Both language files were renamed to match the new date prefix, along with the linkedin/ sidecar that shares their filename. The URL is unchanged in both languages, since the slug drops the date prefix.
v0.5.6
- Fixed a red build that was blocking the site from deploying. "The Agent Was Wrong All Tuesday" was published from the CMS on September 11 with both languages correctly set to draft: false, but the Spanish mirror had never been refreshed after the English rewrite and still carried two [AMIT: confirm] markers. The build guard caught them and failed, as designed, so nothing shipped.
- Restored the English essay's ending. The CMS edit had truncated the body from 5,301 to 2,618 bytes, cutting it off mid-sentence after "And" and losing three sections entirely: "Someone has to notice", "What it costs when you're slow", and "The part I keep coming back to". The sections were recovered from the pre-edit commit and reworked into the shorter-paragraph rhythm of the rewrite.
- Re-mirrored the Spanish essay in full against the rewritten English. Both [AMIT: confirm] paragraphs and the unverified claims they guarded (the specific incident details and how often the manual override was used in the first months) were cut from both languages rather than guessed at.
Rejected this release:
- The specific incident behind the essay (what the bug was, how long it ran before someone noticed, what was pushed through the override) and the frequency of override use in the first months. Both were still open [AMIT: confirm] questions in voice/story-bank.md and could not be verified, so the claims were removed rather than published.
v0.5.5
- Site audit: checked all 69 external links across en/, es/ and data/ — all reachable. openai.com and linkedin.com/in/amitkvint returned 403/999 to both curl and WebFetch, consistent with those vendors' known bot-blocking of automated fetches rather than a dead page, so neither link was changed. Spot-checked the two most recently touched mirror pairs (the field-notes weeks of August 31 and September 7, and the ATLAS.ti qualitative-coding essay) for parity drift — none found. All 8 reference entries in data/reference/ fall within the 60-day freshness window, so none needed re-verification this run.
- SEO: trimmed three Spanish meta descriptions that exceeded the 160-character limit — es/field-notes/2026-09-07.md, es/services.md, and es/writing/2026-09-02-the-queue-looked-fine-on-monday.md — without dropping or softening any claim. Confirmed sitemap.xml and both feed.xml files are well-formed XML, and that no two pages share a title or description.
Verified against: https://arxiv.org/abs/2608.10224 · https://automattic.com/2026/08/31/customer-support-ai-deflection/ · https://commission.europa.eu/news-and-media/news/safer-and-more-transparent-ai-2026-08-02_en · https://crowdin.com/ · https://crowdin.com/blog · https://crowdin.com/blog/whats-new-at-crowdin-august-2026 · https://decagon.ai/ · https://decagon.ai/blog/audio-native-semantic-speaker-change-detection · https://decagon.ai/blog/campaign-composer · https://decagon.ai/blog/gpu-efficient-inference-serving-stack · https://decagon.ai/blog/live-web-search-with-perplexity · https://docs.anthropic.com/ · https://finance.yahoo.com/markets/stocks/articles/salesforce-delivers-record-second-quarter-200100684.html · https://github.com/amitkvint · https://github.com/amitkvint/atlasti-guided-coding · https://global.chinadaily.com.cn/a/202608/19/WS6a8506aaa3106bc57421c3a1.html · https://gurusup.com/ · https://help.salesforce.com/ · https://kvint.es · https://lilt.com/ · https://lilt.com/blog · https://lilt.com/blog/september-2026-monthly-product-release · https://lokalise.com/ · https://lokalise.com/blog/ · https://lokalise.com/blog/ai-translation-quality-evaluation/ · https://openai.com/ · https://phrase.com/ · https://phrase.com/news/ · https://platform.claude.com/docs/en/cli-sdks-libraries/cli/apply · https://platform.claude.com/docs/en/release-notes/overview · https://platform.openai.com/docs/ · https://profiles.wordpress.org/webifish/ · https://salesforcetime.com/2026/08/27/claudeforce-salesforce-meets-claude/ · https://sierra.ai/ · https://sierra.ai/blog/hyper-t-bench-evaluating-agents-that-build-agents · https://sierra.ai/blog/navigating-ivr-systems · https://sierra.ai/blog/sierra-launches-in-korea · https://support.zendesk.com/hc/en-us/articles/11068070409882-What-s-new-in-Zendesk-August-2026 · https://support.zendesk.com/hc/en-us/sections/4405298877338 · https://unbabel.com/ · https://unbabel.com/blog/ · https://wordpress.tv/speakers/amit-kvint/ · https://wpml.org/ · https://wpml.org/category/announcements/ · https://www.anthropic.com/ · https://www.customerexperiencedive.com/news/customers-still-want-access-to-a-human-in-customer-service/828755/ · https://www.deepl.com/ · https://www.deepl.com/en/blog · https://www.digitaltoday.co.kr/en/view/97987/kt-begins-rebuild-of-woori-bank-ai-chatbot-consultation-bot · https://www.freshworks.com/freddy-ai/ · https://www.freshworks.com/whats-new/ · https://www.hubspot.com/new · https://www.hubspot.com/products/artificial-intelligence · https://www.intercom.com/changes · https://www.intercom.com/changes/en · https://www.intercom.com/changes/en/156078-set-more-complex-series-live-reliably · https://www.intercom.com/changes/en/156079-keep-your-integrations-in-sync-with-translated-replies · https://www.intercom.com/changes/en/156211-track-data-connector-health-over-time · https://www.intercom.com/changes/en/156596-see-how-long-tickets-spend-with-every-team-that-touches-them · https://www.intercom.com/changes/en/156649-fin-can-now-query-your-snowflake-data-warehouse · https://www.intercom.com/fin · https://www.linkedin.com/in/amitkvint/ · https://www.salesforce.com/agentforce/ · https://www.salesforce.com/news/stories/salesforce-simplifies-editions-2026/ · https://www.salesforce.com/news/stories/winter-2027-product-release-announcement/ · https://www.smartcustomerservice.com/Columns/Vendor-Views/Stop-Counting-Deflections-Why-Resolution-Rate-Is-the-Metric-AI-Customer-Service-Actually-Needs-175977.aspx · https://www.smartling.com/ · https://www.smartling.com/blog/ · https://www.zendesk.com/service/ai/
v0.5.4
- Field notes: published the week of September 7, 2026 digest - Sierra's open-sourced hyper-τ-bench and the gap it measures between a model working alone (23.9% of held-out tasks) and the same model paired with an engineer (82.2%), Fin querying a Snowflake warehouse through an MCP connector, Intercom's ticket assignments dataset and its connector-health charts, and Lilt's In-Context Preview and Context Database.
- Reference refresh: Decagon's summary corrected in data/reference/ai-support-agents.json. Its site describes agents working across voice, chat and email, configured with natural-language Agent Operating Procedures; the previous summary called the product standalone over existing helpdesks, which the site does not claim. Lokalise, Smartling, Crowdin and GuruSup re-checked against their sources and re-dated with no change to their summaries.
Rejected this release:
- Anthropic's ant apply command (CLI 1.30.0, Sept 3) - read and verified this session, then dropped as a duplicate: the week of August 31 digest already covers it.
- Zendesk's September 2026 what's-new page - real-time AI translation extended to live chat and messaging, and AI agent cloning that copies procedures, actions and knowledge sources. The capabilities were read on the page, but it carries no publication date on any individual item, so there is no way to confirm the changes landed this week rather than earlier in the month.
- Unbabel - the blog could not be reached; unbabel.com/blog redirects to a staging hostname that does not serve the archive.
- OpenAI's ChatGPT release notes and the Freshdesk release notes - both returned HTTP errors this session (403 and 404), so nothing from either vendor is reported.
Verified against: https://sierra.ai/blog · https://sierra.ai/blog/hyper-t-bench-evaluating-agents-that-build-agents · https://www.intercom.com/changes/en · https://www.intercom.com/changes/en/156649-fin-can-now-query-your-snowflake-data-warehouse · https://www.intercom.com/changes/en/156596-see-how-long-tickets-spend-with-every-team-that-touches-them · https://www.intercom.com/changes/en/156211-track-data-connector-health-over-time · https://lilt.com/blog · https://lilt.com/blog/september-2026-monthly-product-release · https://platform.claude.com/docs/en/release-notes/overview · https://lokalise.com/blog/ · https://www.smartling.com/blog/ · https://crowdin.com/blog · https://decagon.ai/ · https://gurusup.com/ · https://www.deepl.com/en/blog · https://wpml.org/blog/ · https://www.salesforce.com/news/ · https://support.zendesk.com/hc/en-us/articles/11180657600026-What-s-new-in-Zendesk-September-2026
v0.5.3
- Field notes: published the week of August 31, 2026 digest - Anthropic's ant apply command, Salesforce's Winter '27 editions folding Agentforce into every tier, Crowdin's August recap, a Lokalise post on AI translation quality, and two Decagon engineering posts.
- Reference refresh: verified Decagon, GuruSup, Lokalise, Smartling, and Crowdin in data/reference/ai-support-agents.json and data/reference/ai-localization.json against their vendor pages; no summary corrections needed.
Verified against: https://platform.claude.com/docs/en/release-notes/overview · https://platform.claude.com/docs/en/cli-sdks-libraries/cli/apply · https://www.salesforce.com/news/all-news/ · https://www.salesforce.com/news/stories/salesforce-simplifies-editions-2026/ · https://crowdin.com/blog · https://crowdin.com/blog/whats-new-at-crowdin-august-2026 · https://lokalise.com/blog/ · https://lokalise.com/blog/ai-translation-quality-evaluation/ · https://decagon.ai/blog · https://decagon.ai/blog/gpu-efficient-inference-serving-stack · https://decagon.ai/blog/audio-native-semantic-speaker-change-detection · https://decagon.ai/ · https://gurusup.com/ · https://gurusup.com/about · https://www.smartling.com/blog/
v0.5.2
- Mirror sync: re-mirrored the ATLAS.ti qualitative-coding essay's Spanish version to match a wording change in the English original (the research work is described as happening on weekends).
v0.5.1
- Services removed from the site navigation in both languages. The page itself (/services/, /es/services/) stays published and reachable by direct link, in the sitemap and with its Service JSON-LD.
v0.5.0
- New page: Academic services (/services/, /es/services/), in the navigation before About. What Amit offers research teams: interview coding in ATLAS.ti in Hebrew, English and Spanish, codebook development, AI-assisted coding with a human in the loop, cross-language projects and workshops; a pilot-first process and a plain statement of how data is handled. Service JSON-LD; the parity check now requires the page in both languages.
v0.4.0
- New essay published: "I Taught Claude to Code My Interviews With Me, Not For Me: AI-Assisted Qualitative Coding in ATLAS.ti", with its Spanish mirror. Amit wrote the text; the draft was prepared in the repo, reviewed and published from the CMS.
- The CMS publish left the draft's LinkedIn HTML comment in the English file, so the build guard refused to deploy; the comment was removed. This release is the deploy.
- The Claude skill described in the essay is now public at github.com/amitkvint/atlasti-guided-coding (MIT) and linked from the essay's setup list.
- Tag archive pages: /tags/ lists every tag with its count, /tags/<tag>/ lists the essays and field notes carrying it, in both languages with reciprocal hreflang. Tag pills on lists and posts now link there. New build guard: tags must be lowercase slugs and a mirror pair must carry identical tags.
- Every essay now opens with a contact line under the title ("Want to talk about this essay? Email me:") showing the address with a copy button and a LinkedIn fallback. A first version used a mailto link, which fails for webmail users, and was replaced the same day.
- Each essay's tags are emitted as the keywords property of its BlogPosting JSON-LD.
Verified against: https://atlasti.com/mcp-release · https://github.com/amitkvint/atlasti-guided-coding
v0.3.4
- Weekly audit: fixed two dead reference-source links. Phrase's source now points to its newsroom (phrase.com/news/) instead of the 404'd /releases/; WPML's now points to its announcements category instead of the 404'd /news/.
- Fixed a dead source link on the Freshworks Freddy AI entry (freshdesk/updates/ 404'd); it now points to freshworks.com/whats-new/, and the summary was broadened to cover the Freshworks suite rather than Freshdesk alone.
- Re-verified the DeepL, Unbabel, and Lilt entries in the AI-localization reference guide against their current blogs; summaries held up, verified dates updated.
- Shortened three Spanish meta descriptions (about page, the WPML essay, and the Sept 3 field notes) that exceeded the 160-character SEO limit; also dropped an em dash from the about-page description per Kvint style.
- Spot-checked three recently-changed EN/ES mirror pairs for drift; none found.
Verified against: https://phrase.com/news/ · https://wpml.org/category/announcements/ · https://www.freshworks.com/whats-new/ · https://www.deepl.com/en/blog · https://unbabel.com/blog/ · https://lilt.com/blog
v0.3.3
- New essay published: "WPML Doesn't Work With My Shop". Amit rewrote the draft in the CMS and published it; the Spanish mirror was retranslated from his final text.
- The CMS publish had flipped both languages to published while the Spanish still carried two review markers, so the build guard refused to deploy. The markers were removed with the resync; this release is the deploy.
v0.3.2
- Published the field notes digest for the week of August 27, 2026: Claude Fable 5.1 and Mythos 5.1, Salesforce's Winter '27 release, Intercom's translated-replies webhooks and Series activation fix, Decagon's Campaign Composer and live web search, and KT's Woori Bank chatbot rebuild.
- Reference refresh (5 entries in data/reference/ai-support-agents.json): Anthropic Claude (Agent SDK), OpenAI (Agents platform), Freshworks Freddy AI, and Salesforce Agentforce re-verified with no change to their summaries. HubSpot Breeze corrected: HubSpot's AI agent suite is now unified under the Agent Hub brand with a customer-facing Customer Agent; Breeze is the assistant that helps build and manage them, not the umbrella brand the old summary implied.
Verified against: https://platform.claude.com/docs/en/release-notes/overview · https://code.claude.com/docs/en/agent-sdk/overview · https://developers.openai.com/api/docs/overview · https://www.freshworks.com/freddy-ai/ · https://help.salesforce.com/ · https://www.hubspot.com/products/artificial-intelligence · https://www.salesforce.com/news/stories/winter-2027-product-release-announcement/ · https://www.intercom.com/changes/en/156079-keep-your-integrations-in-sync-with-translated-replies · https://www.intercom.com/changes/en/156078-set-more-complex-series-live-reliably · https://decagon.ai/blog/campaign-composer · https://decagon.ai/blog/live-web-search-with-perplexity · https://www.digitaltoday.co.kr/en/view/97987/kt-begins-rebuild-of-woori-bank-ai-chatbot-consultation-bot
v0.3.1
- Amit rewrote all three essays in English via the CMS (softer, more conversational register; era-accurate numbers; 'heavily involved' phrasing). Spanish mirrors fully resynced to the new texts.
- Fixed a CMS editing artifact: saves had re-flagged two published essays as drafts (unpublishing them) and duplicated the title as an in-body heading. Publish state restored; headings deduplicated.
- Voice rules updated with editing patterns observed in tonight's rewrite.
v0.3.0
- Homepage redesign, approved by Amit: a simulated AI-support chat window joins the hero (customer message, AI-labeled agent reply, and a visible 'Talk to a human' handoff with a 'Resolved · 4 min' stamp) — localized in both languages and theme-aware.
- Latest writing and Latest field notes now sit in two columns on wide screens: essay cards on the left, a compact field-notes list on the right; single column on small screens.
v0.2.1
- First field-notes digest published (week of August 25, 2026): Intercom Fin cross-session memory, Zendesk bundling copilot triage, Salesforce-in-Claude, Sierra outbound IVR navigation and Korea launch, EU AI Act Article 50 in force, Gartner's 87% human-access finding, LinkedIn's production A/B numbers. EN+ES in the same release.
- Reference guide 'AI agents for customer support': Intercom Fin, Sierra, and Zendesk AI agents entries verified against their sources; Intercom and Sierra summaries updated with August capabilities.
Verified against: https://www.intercom.com/changes/en · https://support.zendesk.com/hc/en-us/articles/11068070409882-What-s-new-in-Zendesk-August-2026 · https://support.zendesk.com/hc/en-us/sections/4405298877338 · https://salesforcetime.com/2026/08/27/claudeforce-salesforce-meets-claude/ · https://finance.yahoo.com/markets/stocks/articles/salesforce-delivers-record-second-quarter-200100684.html · https://sierra.ai/blog/navigating-ivr-systems · https://sierra.ai/blog/sierra-launches-in-korea · https://commission.europa.eu/news-and-media/news/safer-and-more-transparent-ai-2026-08-02_en · https://www.customerexperiencedive.com/news/customers-still-want-access-to-a-human-in-customer-service/828755/ · https://arxiv.org/abs/2608.10224
v0.2.0
- Writing section goes live with the first three essays, EN+ES, approved by Amit: 'Measure the Queue, Not the Bot', 'Your Ticket Queue Is a Defect Log', 'Scaling Support Without Hiring'.
- The three external sources cited in 'Measure the Queue' were each fetched and verified against their claims before linking (Automattic, Aug 31; Smart Customer Service, Aug 3; China Daily, Aug 19).
- Site launched: amitkvint.com connected via Cloudflare Workers static assets; launch gate (3 essays + About + reference guide) met.
- Post URLs drop the filename date prefix; duplicate-slug check added to the build.
Verified against: https://automattic.com/2026/08/31/customer-support-ai-deflection/ · https://www.smartcustomerservice.com/Columns/Vendor-Views/Stop-Counting-Deflections-Why-Resolution-Rate-Is-the-Metric-AI-Customer-Service-Actually-Needs-175977.aspx · https://global.chinadaily.com.cn/a/202608/19/WS6a8506aaa3106bc57421c3a1.html
v0.1.0
- Site scaffold built: dependency-free Python generator, English at /, Spanish mirrored at /es/ with reciprocal hreflang, RSS feeds, sitemap, JSON-LD.
- Sections created: Writing, Field notes, Reference, Changelog, About.
- First versions of the two reference guides seeded — every entry deliberately marked 'not yet verified' until the verification pipeline checks it against its source.
- Publishing contract encoded in the repository: essays require a human merge; the automated lane (reference, field notes, mirroring, maintenance) merges itself only when all checks pass.
- CI checks: build, EN/ES parity, internal links, data validation.