Skip to content

docs(blog): the trust-layer essay on context rot and ADD - #205

Merged
TinDang97 merged 1 commit into
mainfrom
docs/blog-trust-layer
Aug 13, 2026
Merged

docs(blog): the trust-layer essay on context rot and ADD#205
TinDang97 merged 1 commit into
mainfrom
docs/blog-trust-layer

Conversation

@TinDang97

Copy link
Copy Markdown
Collaborator

What

Adds blog/add-trust-layer-agentic.html — an editorial-scroll HTML essay for
engineering leaders evaluating ADD. Same convention as the existing
blog/add-25-vs-30-ledger.html: an artifact-style fragment, self-contained,
no CDN and no external fonts.

The piece runs a reason → resolve loop in eight acts:

Act Beat
01 One call vs. an agentic loop — why one response was never enough
02 Context rot: the n² curve, plus Anthropic's four long-running-agent failure modes
03 The four context-engineering strategies — and the authority hole each leaves
04 ADD's answer: the task node, two memory tiers, the five living specs
05 The component digest — every piece of ADD, one line each
06 The transfer-money walkthrough, end to end
07 Capability matrix + the benchmark
08 Adoption

Where the claims come from

Everything factual is traceable to this repo or to Anthropic's published
material — nothing is invented to make the argument land:

  • Failure modes are quoted from Anthropic's Effective harnesses for
    long-running agents
    . Mode 03 is paraphrased and says so on the page.
  • Reveal panels under each failure mode are labelled
    illustrative reconstruction — they show the shape of the behaviour, and
    they are not presented as captured logs.
  • The walkthrough is the canonical transfer example from
    GETTING-STARTED.md, including the verbatim refusal string
    (cannot freeze \transfer` — these (dimension, surface) pairs are unswept: who:S1`).
  • Fig. 4 quotes real ## Deltas entries out of .add/specs/.
  • Fig. 7 plots only the two measured points from README.md
    (coverage .92 → .75 in one continued conversation; ADD at 1.0 on fresh
    sessions). The unmeasured path is dashed and the table view says
    not plotted rather than filling in a curve.

Agentic turn counts and token deltas in act 01 are captioned
illustrative of the shape, not a measurement.

Craft notes

  • Chart colours were run through the dataviz palette validator, not
    eyeballed, in both modes. Light #A8382F / #3E9E7C (deutan ΔE 13.4,
    normal-vision 27.1); dark #C45448 / #45AA88 (deutan ΔE 9.5, normal 25.2).
  • Theme-aware across all three states — bare :root, the guarded
    prefers-color-scheme block, and [data-theme].
  • Keyboard operable: the act-03 tablist takes arrow keys / Home / End; the
    act-02 reveals open on focus as well as hover.
  • Honours prefers-reduced-motion; every chart has a table view; status is
    never encoded by colour alone.
  • No innerHTML anywhere in the inline script.

Note for the reviewer

mkdocs.yml sets docs_dir: add-method/docs, so root blog/ is not built
into the published site
. This lands the piece under version control and up
for review; it does not put it on pilotspace.github.io. Wiring it into the
site (a copy under add-method/docs/announcements/, plus nav) is a separate
call I deliberately left to you, since it changes what ships publicly.

Preview while reviewing: open the file directly in a browser, or read it as a
published artifact at
https://claude.ai/code/artifact/bdef431b-6d8e-4224-8456-27c0f132688f

An editorial-scroll HTML piece aimed at engineering leaders evaluating ADD.
It runs a reason -> resolve loop in eight acts: why one LLM response was never
enough, how context rot presents in production, why the four documented
context-engineering strategies leave an authority hole, and what ADD's three
beats put in that hole.

Grounded throughout in this repo and in Anthropic's published material:

- Anthropic's four long-running-agent failure modes are quoted (mode 03 is
  paraphrased and labelled as such); each row reveals an illustrative
  transcript reconstruction, marked as a reconstruction, not a captured log.
- The four context-engineering strategies are tabbed, each with an SVG that
  draws the same thing: the missing authority slot, dashed.
- The walkthrough is the canonical transfer-money example from
  GETTING-STARTED.md, including the verbatim refusal string.
- Fig. 4 quotes real Deltas from .add/specs/.
- The benchmark chart plots only the two measured points from README.md; the
  unmeasured path is dashed and the table view reads "not plotted".

Chart colors were validated with the dataviz palette checker rather than
eyeballed, in both light and dark mode. The page is self-contained (no CDN,
no external fonts), theme-aware across all three theme states, keyboard
operable, honours prefers-reduced-motion, and carries a table view under
every chart. No innerHTML anywhere.

Follows the blog/add-25-vs-30-ledger.html convention: an artifact-style
fragment, not published by mkdocs (docs_dir is add-method/docs).

author: Tin Dang
@TinDang97
TinDang97 merged commit 503f3f9 into main Aug 13, 2026
8 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant