All videosContext engineering2:13

Extensions, delegation and controls

Progressive loading limits context. It neither isolates the process nor grants new authorisation.

Links containing ?t= open the video at a specific second.

Video summary

The ideas to retain

01

1. Four primitives, four different questions

There is no universal semantics for skill, plugin, subagent, or hook across products. Start with the function each one performs in a specific implementation.

02

2. A skill is reusable expertise, not a security boundary

Anthropic Agent Skills provides a current, concrete example. A skill is a directory containing SKILL.md, instructions, and optionally code and reference material. The mechanism uses…

03

3. Progressive disclosure lowers initial context cost, not activation cost to zero

A skill must be discoverable before it can be selected. That creates two different costs:

Key moments

Jump directly to a section

  1. Load a skill in stages
  2. A package is not a sandbox
  3. Delegate a precise contract
  4. Another window is not another environment
  5. Block before the effect
  6. Measure success and isolation separately
Read the reviewed transcript

This video has no narration. This transcript reproduces its on-screen text; it does not invent a spoken track.

Load a skill in stages

The catalogue shows a skill description without loading all its instructions.

When needed, its procedure is loaded, followed by relevant references.

Progressive loading limits context. It neither isolates the process nor grants new authorisation.

A package is not a sandbox

A plugin distributes a skill, a hook and an integration.

Those components may run on different surfaces with different permissions.

Installing the package does not prove isolation. Inspect each component and its environment.

Delegate a precise contract

The parent needs a review of diff A, not a copy of the entire conversation.

The envelope fixes the goal, allowed files, tools, budget and output format.

The specialist returns findings and evidence for A. A summary without identity cannot close the review.

Another window is not another environment

Two agents have separate model contexts but share the same filesystem.

Separating files, tools and network needs additional controls. The word subagent does not imply them.

Evaluation checks each boundary and prohibited accesses, as well as task success.

Block before the effect

A preventive check decides before a write executes. Denial means no effect.

A post-action hook observes a completed write. Running later does not undo it.

Ordering is part of the contract. Test the event, failure and actual decision.

Measure success and isolation separately

A specialist may solve the task while reading a prohibited file.

Another run respects isolation but returns an incorrect result. These are different failures.

Delivery needs task and boundary evidence for the same run, not a single PASS label.