Video summary

Ten months ago I said AI changed. It just happened again.

Main summary

Key takeaways

News and Commentary

Overview

The speaker argues that AI progress has crossed another “boundary,” similar to a major shift they described about 10 months earlier (around November 2025). They use the idea of KPG/KT geological boundaries: the change may not be obvious at first, but becomes clear months later when widespread behavior shifts.

Main arguments and timeline of change

Earlier boundary (around November 2025): from task-by-task to multi-step execution

Before this shift, the workflow often worked like a tight loop:

  • A user asks for a small task
  • The AI returns an answer
  • The user reviews it
  • The user asks for the next change

When users tried to have the AI do multiple tasks at once, it often failed due to issues such as:

  • lost context
  • “collapse” in coherence
  • hallucinations
  • erroneous changes

As a result, people developed scaffolding approaches, for example:

  • spec-driven development
  • file-based state
  • rule sets

What changed around November 2025

Around November 2025, models became capable of handling multiple units of work with:

  • dependencies and priorities
  • tracking over time
  • internal checking/fixing
  • a more complex internal loop

This impact wasn’t immediate for most users. However:

  • By March, more people began benefiting broadly
  • By May / early summer, users could request complex deliverables in a single prompt that the system could execute for an hour or two, returning “complete work”

Current boundary (“this time”): destination/goals instead of just tasks

For roughly the past year, users mostly described:

  • the work itself (task lists, steps, or long single-session requests)

The new capability is that users can describe the destination—the orientation toward work:

  • what success looks like
  • how the system knows the work is done

The system then orchestrates the work needed to reach that goal.

Internal structuring (“factories”)

The speaker notes that the system often creates its own internal structure (including “agent building agents”) to:

  • design roles
  • set up loops
  • run checks and rules

Evidence in small features, plus broader capability

They argue this can be seen in smaller features such as “slashgoal” in tools, but also claims it can happen without that explicit feature. The agent can:

  • plan
  • evaluate progress
  • write constraints
  • continue until the goal is met

Why they emphasize “new models” rather than only better user techniques

The speaker attributes the shift to new major model releases, not just improved prompting patterns.

They mention models such as:

  • Fable
  • Astra

Their claim is that new models introduce new connectivities/associations, enabling capabilities like:

  • staying on the right track
  • reliably determining when work is actually finished

They also describe personal experience:

  • With older models, they tried building complex agent systems, but it often took days of scaffolding and still failed frequently.
  • With newer models, they can “throw away” much of that scaffolding because orchestration is handled more inherently.

They note the approach is currently expensive and token-heavy, but:

  • the quality
  • the factory-like orchestration performance

are far better than even a month earlier.

Astra is singled out as exceptionally strong for this kind of work, with Fable also praised.

Implications for non-expert users

Even if most people aren’t building agent systems, the speaker argues that:

  • most users are already experiencing the previous boundary
  • simple multi-sentence instructions already produce near-expected outcomes

They predict that within 6 months (and by Christmas/January for many), most users will naturally start asking in the new destination/orientation way—without needing specialized thinking.

However, there’s a learning curve:

  • People typically think in task-level steps (finish A before B in a single thread).
  • The new approach operates more directly at the level of broader goals, which may be unfamiliar and requires practice.

The speaker emphasizes that this reorientation matters because productivity gains scale dramatically once workflows support system-level goals.

What they plan to share next

The speaker says they have been researching for months to map:

  • where these boundaries occur
  • how to build tools to support them

They mention areas of focus such as:

  • rule sets
  • roles
  • visualizations
  • protocols for multi-agent communication
  • reporting

They plan to publish their methodologies and discoveries on the channel for “at least a few months,” and also point out that tools are improving rapidly—citing Cursor releasing a feature that moves closer to these goal-oriented workflows. They encourage viewers to try advanced models with harder challenges.

Presenters or contributors

  • Primary presenter: The unnamed author/speaker of the channel (referred to as “Hey, I’m back…”)

Original video