Context Language Models
Agents organise their own working information to keep track of longer tasks.
0 → 1Earlier → later videosResearch ↗
YOUTUBE INTELLIGENCE · SIX-WEEK REPORT
What’s changing in AI?
New technical developments and changing discussion, drawn from our saved YouTube transcripts.
Report through25 Oct 2026Figures refreshed 31 Oct 2026
Videos in the archive8,188
Saved transcripts5,237
Channels represented829
Entire archive · Totals checked 31 Oct 2026 · Includes videos outside the comparison dates
Earlier six weeks3 Aug 2026–13 Sep 2026992 recorded videos · 570 transcripts analysed
Later six weeks14 Sep 2026–25 Oct 20261,644 recorded videos · 1,187 transcripts analysed
Which topics gained a larger share?
Share of analysed transcripts · Matching videos below each percentage
Hover over a topic, focus it with the keyboard, or tap it to see what it includes.
| Topic | 3 Aug 2026–13 Sep 2026570 transcripts | 14 Sep 2026–25 Oct 20261,187 transcripts |
|---|---|---|
Agent teamsAgent teams Covers multi-agent systems, subagents, multiple agents, teams, swarms, parallel agents and orchestration: organising agent work. Subagents are included in this broad topic. The software factory wording alone does not qualify a video for this row. A video discussing both agent teams and software factories can appear in both rows. |
14.2%81 videos | 12.4%147 videos |
Coordinating agents and swarmsCoordinating agents and swarms Explicit discussion of multi-agent systems, teams, swarms, parallel agents or orchestration: organising how agents divide and coordinate work. Mentioning subagents alone does not qualify for this narrower row. A video may also appear in the broader agent teams category. |
6.3%36 videos | 7.7%91 videos |
Security, approvals and evaluationSecurity, approvals and evaluation Discussions of prompt injection, agent security and identity, observing agent behaviour, evaluating results, human approvals, approval gates, sandboxes and guardrails. These cover how agent actions are monitored, checked or restricted. |
3.3%19 videos | 5.1%61 videos |
Reusable agent skillsReusable agent skills Reusable instructions and procedures packaged as agent skills, including SKILL.md files, custom skills, skill folders, packages, libraries, registries and standards. Also includes progressive disclosure: loading instructions when they are needed. |
6.0%34 videos | 5.1%60 videos |
Hosted and background agentsHosted and background agents Agents described as hosted, cloud-based, remote, asynchronous, persistent, always-on, background or long-running, including durable execution that preserves progress through interruptions. |
4.0%23 videos | 5.0%59 videos |
Voice agents and computer/browser useVoice agents and computer/browser use A combined category covering voice agents and live or real-time voice, plus computer use, browser agents and browser automation. A video can qualify through voice interaction, computer/browser use, or both; it is counted once in this row. |
4.2%24 videos | 4.2%50 videos |
Memory and context managementMemory and context management Persistent and long-term memory, memory banks, knowledge graphs, context engineering and management, Context Language Models, context compaction and context rot. These concern keeping useful information available to an agent. |
1.8%10 videos | 2.2%26 videos |
Software factoriesSoftware factories Explicit discussion using the wording software factory or software factories. These describe workflows linking stages such as planning, coding, testing and review. This row is scored separately from agent teams; general discussion of multiple agents alone does not qualify. A video covering both topics is counted once in each row. |
2.1%12 videos | 1.9%22 videos |
Self-improving agent systemsSelf-improving agent systems Discussion of self-improving or self-learning agents and recursive harness improvement, in connection with agent workflows, prompts, tools or software factories. This measures discussion of improvement methods; a matching video does not establish that a system improves reliably. |
0.5%3 videos | 0.4%5 videos |
Agent teams and software factories are scored separately. 9 videos qualify for both in the earlier period; 9 in the later period.
Coordinating agents and swarms requires explicit discussion of coordination; mentioning subagents alone does not qualify.
Topics overlap, so a video may appear in more than one row. Each video is counted once per matching topic.
How these figures were calculated
The comparison uses publication dates in New Zealand time and English transcripts containing at least 1,000 characters. A topic needs at least three qualifying phrase matches, or a matching title and at least one qualifying phrase in its transcript. Confirmed duplicate excerpts are excluded.
| Collection coverage | Earlier six weeks | Later six weeks |
|---|---|---|
| Recorded videos | 992 | 1,644 |
| Saved transcripts | 571 | 1,198 |
| Transcript availability | 57.6% | 72.9% |
| Transcripts analysed | 570 | 1,187 |
Shares describe discussion within this saved collection. The channel mix and transcript availability changed between periods; the figures do not measure wider industry adoption.
Technical developments to watch
Named developments appearing in the later period
Harness²
Agents test changes to their instructions, tools and methods, and retain useful lessons.
0 → 2Earlier → later videosResearch ↗
Knowledge Catalog
Agents identify gaps in shared business knowledge and propose corrections others can use.
0 → 1Earlier → later videosAnnouncement ↗
Counts show matching analysed videos in the two periods above. Zero means no match in that sample.
ALSO THIS PERIOD
Agent Skills donated to the Agentic AI Foundation ↗ · Announced 22 Oct 2026. The donation is new; reusable skills were already established.
Watch the developments
Video publication dates · New Zealand time