Context Language Models
Agents organise their own working information to keep track of longer tasks.
YOUTUBE INTELLIGENCE · MONTHLY REPORT
Track six months of AI discussion using saved YouTube transcripts, topic counts and source videos.
1 May 2026–31 Oct 2026 · Video publication dates in New Zealand time
FOCUS AREA
Workflow design, automation platforms, triggers, recovery, human approvals and practical business processes.
distinct videos across the workflow categories below. Videos may also appear in other topic families.
2,935 analysed videos = 2,749 matching a displayed topic + 186 other relevant AI videos.
3,098 saved transcripts screened. All headline figures and topic counts use the same saved analysis, refreshed 16 Nov 2026.
Coverage revised 2 Nov 2026. All six months were recalculated with the expanded categories and content checks; earlier report editions used narrower coverage.
Weekly refresh starts Monday at 8:00 am New Zealand time, using saved transcripts and the six latest completed calendar months.
Matching videos · Share of analysed transcripts in brackets · Workflow categories first · Highest totals within each family
Nov 2026 MTD covers saved transcripts through 2026-11-11. Total and headline cards cover the six completed months.
Click a count to see its videos. Hover over a topic, focus it with the keyboard, or tap it to see what it includes.
| Topic | Total2,935 analysed | Nov 2026 MTD414 analysed through 2026-11-11 | Oct 2026933 analysed | Sep 2026563 analysed | Aug 2026363 analysed | Jul 2026335 analysed | Jun 2026375 analysed | May 2026366 analysed |
|---|---|---|---|---|---|---|---|---|
| Agentic workflow automation | ||||||||
Coordinating agents114 channels · Six monthsCoordinating agents Coordinating multiple agents and explicit agent handoffs. |
(11.2%) | (6.5%) | (11.1%) | (9.4%) | (10.5%) | (8.4%) | (15.7%) | (12.6%) |
Automation platforms61 channels · Six monthsAutomation platforms Building AI-enabled workflows in n8n, Make, Zapier, Power Automate, Power Platform or Copilot Studio. |
(5.6%) | (4.1%) | (6.5%) | (4.4%) | (8.0%) | (5.4%) | (5.3%) | (3.3%) |
Hosted and background agents66 channels · Six monthsHosted and background agents Hosting agents, background execution and long-running tasks. |
(5.1%) | (3.9%) | (7.3%) | (5.5%) | (3.0%) | (3.9%) | (3.7%) | (3.6%) |
Agentic workflow design55 channels · Six monthsAgentic workflow design Designing workflows that combine AI decisions, tools and explicit process steps. |
(4.8%) | (3.9%) | (3.1%) | (5.0%) | (5.2%) | (5.1%) | (5.6%) | (7.1%) |
Triggers and scheduling37 channels · Six monthsTriggers and scheduling Starting AI workflows from events, webhooks or schedules. |
(3.7%) | (1.0%) | (2.3%) | (2.8%) | (4.1%) | (6.6%) | (3.5%) | (6.3%) |
Human oversight and approvals53 channels · Six monthsHuman oversight and approvals Human review, approval gates and handoffs in AI workflows. |
(3.3%) | (2.2%) | (4.4%) | (2.8%) | (2.2%) | (2.4%) | (2.7%) | (3.8%) |
Research and information workflows33 channels · Six monthsResearch and information workflows Agents gathering, comparing and synthesising information across research steps. |
(3.1%) | (0.5%) | (1.2%) | (3.0%) | (3.6%) | (2.7%) | (5.9%) | (4.9%) |
Retries and workflow recovery38 channels · Six monthsRetries and workflow recovery Handling failed AI workflow steps, retries, checkpoints and safe resumption. |
(2.1%) | (3.4%) | (2.0%) | (1.6%) | (2.5%) | (1.2%) | (2.9%) | (2.7%) |
Sales and CRM workflows21 channels · Six monthsSales and CRM workflows AI workflows for sales, lead research, outreach and CRM work. |
(1.8%) | (1.0%) | (1.2%) | (1.2%) | (2.5%) | (3.3%) | (1.3%) | (2.5%) |
Customer support workflows29 channels · Six monthsCustomer support workflows AI-assisted support, service agents and ticket handling. |
(1.5%) | (1.4%) | (1.8%) | (1.8%) | (0.8%) | (1.5%) | (0.8%) | (1.6%) |
Marketing and content workflows13 channels · Six monthsMarketing and content workflows AI workflows for marketing, writing, publishing and content production. |
(0.6%) | 0 (0.0%) | (0.2%) | (0.5%) | (1.4%) | (0.6%) | (1.3%) | (0.5%) |
Email, office and reporting workflows11 channels · Six monthsEmail, office and reporting workflows AI workflows for email, calendars, meetings, spreadsheets and reporting. |
(0.6%) | (0.5%) | (0.3%) | (0.4%) | (0.8%) | (1.2%) | (1.1%) | (0.3%) |
Document processing workflows8 channels · Six monthsDocument processing workflows AI workflows for extracting, checking and processing documents and invoices. |
(0.4%) | 0 (0.0%) | (0.3%) | (0.4%) | (0.6%) | (0.3%) | (0.5%) | (0.3%) |
| Building and operating agents | ||||||||
MCP, tools and integrations220 channels · Six monthsMCP, tools and integrations Connecting agents to tools, APIs and services through MCP, tool calling and integration protocols. |
(25.5%) | (18.4%) | (24.9%) | (21.1%) | (24.8%) | (27.5%) | (28.0%) | (30.3%) |
Agent security and permissions221 channels · Six monthsAgent security and permissions Securing AI systems, permissions, prompt injection, isolation and guardrails. |
(17.3%) | (17.9%) | (24.2%) | (14.7%) | (14.3%) | (14.3%) | (13.9%) | (12.8%) |
Evaluation, testing and review185 channels · Six monthsEvaluation, testing and review Evaluating AI outcomes, testing agent work and reviewing AI-generated code. |
(16.2%) | (16.4%) | (20.3%) | (17.2%) | (12.7%) | (13.4%) | (15.7%) | (10.9%) |
Agent teams and subagents126 channels · Six monthsAgent teams and subagents Multiple agents, subagents and teams. Coordination and swarms have their own more specific rows. |
(14.4%) | (9.7%) | (11.7%) | (12.4%) | (18.7%) | (14.3%) | (17.1%) | (17.2%) |
Context management107 channels · Six monthsContext management Selecting and managing the information available during AI work. |
(10.0%) | (8.5%) | (8.7%) | (8.9%) | (9.6%) | (9.6%) | (11.2%) | (14.8%) |
Deployment and production operations114 channels · Six monthsDeployment and production operations Deploying, operating and maintaining AI applications and agent services. |
(9.9%) | (6.3%) | (10.0%) | (9.4%) | (9.6%) | (9.3%) | (11.7%) | (9.8%) |
Retrieval and RAG63 channels · Six monthsRetrieval and RAG Retrieving evidence and enterprise knowledge for AI answers and agent decisions. |
(7.8%) | (2.2%) | (5.1%) | (5.9%) | (5.5%) | (11.9%) | (13.3%) | (10.7%) |
Reusable agent skills83 channels · Six monthsReusable agent skills Reusable skill files, packages and libraries for agent tasks. |
(7.0%) | (5.1%) | (5.6%) | (6.4%) | (7.2%) | (8.4%) | (7.2%) | (9.8%) |
Agent monitoring and tracing67 channels · Six monthsAgent monitoring and tracing Observing agent behaviour through monitoring, logs, traces and telemetry. |
(5.3%) | (4.6%) | (6.1%) | (5.0%) | (3.9%) | (2.4%) | (8.0%) | (5.2%) |
Agent SDKs and frameworks31 channels · Six monthsAgent SDKs and frameworks Developing agent applications with SDKs, frameworks and agent runtimes. |
(4.0%) | (1.9%) | (3.1%) | (3.2%) | (1.9%) | (4.2%) | (6.4%) | (6.8%) |
Computer use37 channels · Six monthsComputer use Agents interacting with desktop applications and computer interfaces. |
(3.8%) | (2.2%) | (2.7%) | (6.9%) | (2.2%) | (5.1%) | (2.9%) | (3.0%) |
Persistent memory36 channels · Six monthsPersistent memory Keeping and reusing agent information across tasks and sessions. |
(3.6%) | (1.0%) | (2.7%) | (3.4%) | (2.2%) | (2.7%) | (5.9%) | (6.6%) |
Software factories21 channels · Six monthsSoftware factories Explicit software-factory approaches to producing software. |
(1.9%) | (1.7%) | (2.9%) | (2.1%) | (2.2%) | (1.2%) | (0.5%) | (0.5%) |
Browser agents and automation24 channels · Six monthsBrowser agents and automation Agents and AI-assisted automation interacting with websites. |
(1.8%) | (0.2%) | (0.9%) | (2.8%) | (1.4%) | (1.2%) | (0.8%) | (4.6%) |
Voice agents19 channels · Six monthsVoice agents Spoken interaction and real-time voice agents. |
(1.2%) | (0.7%) | (0.5%) | (2.0%) | (1.4%) | (1.8%) | (0.5%) | (1.4%) |
Knowledge graphs15 channels · Six monthsKnowledge graphs Representing knowledge and relationships for AI retrieval and reasoning. |
(1.1%) | (0.5%) | (0.2%) | (1.1%) | (0.8%) | (1.5%) | (2.1%) | (2.2%) |
Self-improving agent systems16 channels · Six monthsSelf-improving agent systems Agents and harnesses that evaluate and improve their own working methods. |
(0.9%) | (0.5%) | (0.6%) | (0.2%) | (1.1%) | (0.9%) | (1.9%) | (1.4%) |
Agent swarms7 channels · Six monthsAgent swarms Explicit discussion of agent swarms. |
(0.4%) | (0.5%) | (0.3%) | (1.1%) | (0.3%) | (0.3%) | 0 (0.0%) | 0 (0.0%) |
| AI development and models | ||||||||
AI coding and software delivery529 channels · Six monthsAI coding and software delivery Using AI coding tools to build, change and deliver software, including Claude Code, Codex, Copilot, Lovable and Replit. |
(61.8%) | (63.3%) | (64.4%) | (61.3%) | (62.8%) | (58.8%) | (57.9%) | (61.7%) |
Cost, tokens and performance388 channels · Six monthsCost, tokens and performance AI usage costs, token consumption, inference speed and performance tradeoffs. |
(48.3%) | (42.0%) | (49.1%) | (51.9%) | (45.5%) | (46.3%) | (52.0%) | (41.8%) |
Models and comparisons162 channels · Six monthsModels and comparisons Model releases, selection, capabilities and comparative testing. Video claims are not independently verified. |
(25.0%) | (16.9%) | (22.0%) | (32.5%) | (21.8%) | (29.3%) | (23.2%) | (22.7%) |
Prompts, instructions and harnesses186 channels · Six monthsPrompts, instructions and harnesses Prompts, repository instructions, rules files, hooks and harness design. |
(20.6%) | (17.1%) | (20.7%) | (17.8%) | (22.6%) | (21.2%) | (20.3%) | (23.0%) |
Local and open models93 channels · Six monthsLocal and open models Running or adapting local, self-hosted and open models. |
(9.1%) | (5.3%) | (7.9%) | (7.6%) | (9.4%) | (11.9%) | (12.5%) | (7.7%) |
| Business and creative uses | ||||||||
Business use, adoption and work137 channels · Six monthsBusiness use, adoption and work Using AI in organisations, work, productivity and commercial services. |
(18.4%) | (13.0%) | (17.1%) | (16.3%) | (19.3%) | (19.1%) | (24.3%) | (17.2%) |
AI assistants and workspaces60 channels · Six monthsAI assistants and workspaces Personal AI assistants, agent workspaces and AI workforces that organise or carry out tasks. |
(8.7%) | (3.4%) | (5.9%) | (5.7%) | (9.6%) | (9.9%) | (12.8%) | (14.5%) |
Creative and multimodal AI45 channels · Six monthsCreative and multimodal AI AI-assisted image, video, audio and multimodal creative work. |
(7.7%) | (1.9%) | (5.5%) | (7.6%) | (9.6%) | (9.3%) | (8.3%) | (9.6%) |
| Total topic matches | (375.6%) | (289.4%) | (364.8%) | (363.4%) | (365.6%) | (377.9%) | (408.8%) | (395.6%) |
| Distinct matching videos | 2,749 | 363 | 877 | 538 | 340 | 309 | 350 | 335 |
| Other relevant AI discussion | ||||||||
Topics overlap, so a video may appear in more than one row. Each video is counted once per matching topic. The Total topic matches row counts overlapping matches; its percentage may exceed 100%. The next two rows count distinct videos and add up to the analysed total.
Each percentage uses that column’s analysed transcripts; the Total column uses all six months combined.
Entire archive: 9,457 videos · 6,350 saved transcripts · 1,486 channels. Checked 16 Nov 2026. These archive totals include videos outside the reporting dates.
For 1 May 2026–31 Oct 2026: 4,772 recorded videos − 4 confirmed duplicate excerpts − 1,633 without an available non-empty transcript − 37 transcripts shorter than 1,000 characters = 3,098 transcripts screened.
After content screening: 121 fail the English text check; 5 need language review; 13 have no qualifying AI evidence; 24 need AI relevance review. The remaining 2,935 videos form the topic-percentage denominator.
Every screened video belongs to exactly one of the categories below. Click a count to inspect the videos; topic rows overlap within the matched category. Channel counts use public IDs, with names as a fallback.
| Screening result | All six months | Oct 2026 | Sep 2026 | Aug 2026 | Jul 2026 | Jun 2026 | May 2026 |
|---|---|---|---|---|---|---|---|
| Matches displayed topics | |||||||
| Other relevant AI discussion | |||||||
| Outside AI scope | |||||||
| AI relevance needs review | |||||||
| Non-English text | |||||||
| Language needs review | 0 | 0 | 0 | 0 |
| Collection coverage | Oct 2026 | Sep 2026 | Aug 2026 | Jul 2026 | Jun 2026 | May 2026 |
|---|---|---|---|---|---|---|
| Videos after duplicate exclusion | 1,252 | 964 | 663 | 600 | 606 | 683 |
| Saved transcripts | 1,006 | 587 | 378 | 361 | 398 | 405 |
| Transcript availability | 80.4% | 60.9% | 57.0% | 60.2% | 65.7% | 59.3% |
| Transcripts analysed | 933 | 563 | 363 | 335 | 375 | 366 |
The report uses the six latest completed calendar months and New Zealand publication dates. Saved transcripts must contain at least 1,000 characters. Confirmed duplicate excerpts are excluded.
Language checks examine the transcript text, including non-Latin scripts and common English and other-language words. Long transcripts use a bounded beginning/middle/end sample. Stored language labels alone do not establish English. Mixed or uncertain language is held for review. This conservative automated check can still make mistakes.
AI relevance requires a qualifying title plus supporting transcript evidence, or repeated AI evidence in separated passages. No qualifying evidence is marked outside scope; a single weak signal is held for review. Neither category contributes to topic percentages.
Topics need a matching title plus supporting transcript wording, or two qualifying passages at least 160 characters apart. Ambiguous terms require nearby AI or workflow context. Definitions include natural wording such as evals, evaluations, monitoring and tracing; they do not require one exact phrase. These are automated discussion indicators, not independently verified subject labels. Other relevant AI discussion remains visible for future category review.
Shares describe discussion within this saved collection. The channel mix and transcript availability changed between periods; the figures do not measure wider industry adoption.
Manually reviewed · Last reviewed 31 Oct 2026
Editorial comparison: 3 Aug 2026–13 Sep 2026 → 14 Sep 2026–25 Oct 2026. This section is updated after manual review.
Agents organise their own working information to keep track of longer tasks.
Agents test changes to their instructions, tools and methods, and retain useful lessons.
Agents identify gaps in shared business knowledge and propose corrections others can use.
Counts refer to the editorial comparison dates shown in this section. Zero means no match in that sample.
Agent Skills donated to the Agentic AI Foundation ↗ · Announced 22 Oct 2026. The donation is new; reusable skills were already established.
Video publication dates · New Zealand time
EDITORIAL UPDATE · 17 NOV 2026
What they make possible · How mature they are
These developments point towards agents retaining useful knowledge, coordinating different models, adapting how they work and completing longer tasks. The evidence spans research and available tools; it does not establish dependable general autonomy.
Background: the initial scan covered 1,194 videos with saved transcripts published 13 September–25 October 2026, with passages examined from 18 selected transcripts. The initial edition’s research and announcement links were checked on 2 November. This manually reviewed overview is separate from the automatically refreshed monthly figures above.
Latest review: 556 newly available saved transcripts from 354 channels were screened through 16 November 2026 at 8:49 pm New Zealand time, with focused close checks for proposed changes. Primary sources were checked on 16 Nov 2026. All eight published directions remain; one additional direction is added. The review does not establish industry growth, adoption or independently tested product performance.
Topic dates show when each direction first appeared in this section and when its content, sources or maturity assessment last changed. All dates use New Zealand time.
Added this review
Agents can execute code and tools in separate environments while controls outside those environments restrict access to files, network destinations and services. Google’s 5 November announcement of Agent Substrate 1.0 and general availability on GKE provides one example. devvm applies project-scoped virtual machines and credential mediation to local coding work. These mechanisms can reduce what a compromised agent can reach.
Available infrastructure and developer tooling, documented rather than independently tested. Sandboxing predates this review. It does not guarantee correct work or remove risks through permitted files and network routes. This is newly featured editorial coverage supported by an availability milestone.
Google availability announcement · 5 Nov ↗devvm safety documentation ↗Google Cloud Tech · 10 Nov NZ ↗Alex Hitt · 9 Nov NZ ↗
Every previously published topic remains. Model coordination moves into this group with its copy and confirmed dates retained.
Retained
Software can choose one model to do a task, ask another to review its work, or escalate a difficult task to a stronger model. GitHub made Copilot Fusion generally available on 28 October, extending its earlier HydraFusion preview. Model coordination becomes part of the coding tool rather than something the developer manages by hand.
Available product. The release is new; multi-model coordination is not. Cost and quality depend on the task and settings, and changes still need review.
Retained
Harness² explores agents changing their tools, instructions and memory arrangements while keeping the underlying model fixed. This could help a system adapt to unfamiliar tasks.
Research stage. Benchmark improvements do not establish dependable general autonomy.
Updated this review
Context Language Models research lets a model edit the information it carries forward. Context Mode Engine instead keeps bulky tool output in a searchable external index. A third approach maintains stored notes between sessions: Anthropic’s 6 November research describes revisiting session logs to correct and organise memory. Its Dreams feature produces a separate output store that can be reviewed.
Research, research preview and available developer tooling. These are different mechanisms. Reorganised notes can still be wrong; they do not guarantee complete memory or change the model’s weights.
Context Language Models ↗Context Mode Engine ↗Memory research · 6 Nov ↗Dreams · research preview ↗Beau Johnson · 9 Nov NZ ↗
Updated this review
Google’s Knowledge Catalog preview analyses agent and user interactions to find gaps in business definitions and propose corrections. A reviewed correction could then help several agents use the same business knowledge. A Google demonstration published on 7 November New Zealand time shows agents consulting shared semantic definitions when answering business questions.
Product preview. Publication can be manual or use configured confidence rules; sensitive changes still need review. The controlled demonstration does not establish accuracy or savings across other organisations.
Google announcement · 22 Oct ↗Google Cloud Tech · 23 Oct NZ ↗Semantic-model demonstration · 7 Nov NZ ↗
Updated this review
Reusable task instructions can be shared across compatible agents. Anthropic announced the transfer of the Agent Skills standard to the Agentic AI Foundation on 22 October. Microsoft’s Copilot Studio example makes this practical: its Microsoft Copilot harness can import Markdown or SKILL.md packages with supporting resources, then load the relevant instructions when needed.
Open standard, with tools at different stages. The cited Copilot Studio agent-creation experience is preview. Portable instructions do not guarantee identical results, tools or permissions. Imported instructions and scripts need review.
Standards announcement · 22 Oct ↗Microsoft skills documentation ↗Import requirements ↗Microsoft Community Learning · 10 Nov NZ ↗
These three ideas predate the original six-week scan. The collection favours agents and software development, with limited coverage of fields such as robotics and biology. The initial scan identified examples; it did not establish growth relative to an earlier period or adoption across the wider industry.
Historical context, checked 30 October 2026. These are the figures from the original comparison and archive audit. They are retained as a dated snapshot, not the current totals. The monthly analysis above uses a later collection and revised content checks.
The later six weeks contained more collected videos, and transcripts were available for a larger proportion of them.
| Collection measure | 3 Aug–13 Sep 2026 | 14 Sep–25 Oct 2026 |
|---|---|---|
| Videos recorded | 995 | 1,648 |
| Saved transcripts | 571 | 1,179 |
| Percentage with transcripts | 57.4% | 71.5% |
| Channels represented by those transcripts | 178 | 544 |
| Transcripts eligible for analysis | 570 | 1,168 |
The sample grew through broader collection and better transcript coverage. These figures alone cannot tell us whether AI discussion across YouTube grew at the same rate.
Videos are grouped by publication month in New Zealand time. Saved transcripts include all languages and lengths.
| Publication month | Archived videos | Saved transcripts |
|---|---|---|
| June 2025 | 1 | 1 |
| July 2025 | 0 | 0 |
| August 2025 | 0 | 0 |
| September 2025 | 2 | 2 |
| October 2025 | 260 | 145 |
| November 2025 | 494 | 301 |
| December 2025 | 394 | 259 |
| January 2026 | 482 | 286 |
| February 2026 | 580 | 369 |
| March 2026 | 672 | 424 |
| April 2026 | 621 | 399 |
| May 2026 | 683 | 405 |
| June 2026 | 606 | 398 |
| July 2026 | 602 | 361 |
| August 2026 | 663 | 378 |
| September 2026 | 982 | 587 |
| October 2026 · through 30 October | 1,098 | 866 |
| Total at 30 Oct 2026 | 8,140 | 5,181 |
By recorded retrieval month, 16 saved transcripts dated to September 2026 and 5,165 to October 2026. Older videos had therefore been retrieved more recently. Those timestamps describe the saved records at that check, rather than a complete download history.
Publication month is the useful basis for discussing trends. The monthly archive snapshot above runs through 30 October; the original six-week comparison ends on 25 October. These historical figures are manually maintained and do not change with the weekly refresh.