The journal

One subject at a time.
All the way down.

The longer thinking behind the videos: what I believe, what I can show, what I still need to test, and the discussion that followed.

WEEKLY MASTER ARTICLES01 — 11
01
Agent engineering · Days 01–07 · 2026-09-25

Agents need an independent exam

The practical controls that let a fast coding agent produce work you can actually sign off.

02
Engineering judgment · Days 08–14 · 2026-09-25

Model consensus is not engineering evidence

What model debate can reveal, what polished probing cannot supply, and how real experiments turn a design into an engineering decision.

03
Model systems · Days 15–21 · 2026-09-25

Where model control actually lives

What a typed decision can actually control: full-path Jev versus small-model economics, provider-owned loops, scoped tools, context and engineering ownership.

The daily record

Every angle has
room to breathe.

Platform editions are tailored to their audience. Their complete written reads are here now; players appear only for verified public videos.

Instagram
DAY 01 / INSTAGRAM

Do not re-teach stock safety per agent.

Watch & read
LinkedIn
DAY 01 / LINKEDIN

Who writes the agent's exam?

Watch & read
TikTok
DAY 01 / TIKTOK

Make retry safety survive every agent run.

Watch & read
X
DAY 01 / X

Fast agents still need fixed contracts.

Watch & read
Instagram
DAY 02 / INSTAGRAM

The agent cannot grade its own click path.

Watch & read
LinkedIn
DAY 02 / LINKEDIN

Computer use needs a second witness.

Watch & read
TikTok
DAY 02 / TIKTOK

Saved is not the same as survived.

Watch & read
X
DAY 02 / X

A screenshot is not an assertion.

Watch & read
Instagram
DAY 03 / INSTAGRAM

AI budget hides in the attention loop.

Watch & read
LinkedIn
DAY 03 / LINKEDIN

Seat adoption is not AI ROI.

Watch & read
TikTok
DAY 03 / TIKTOK

Five seats can make one manager the queue.

Watch & read
X
DAY 03 / X

The cheap agent run can create an expensive loop.

Watch & read
Instagram
DAY 04 / INSTAGRAM

Let agents move within real boundaries.

Watch & read
LinkedIn
DAY 04 / LINKEDIN

Agentic speed needs observable control.

Watch & read
TikTok
DAY 04 / TIKTOK

The guardrail must travel with the work.

Watch & read
X
DAY 04 / X

Fast agents need upstream controls.

Watch & read
Instagram
DAY 05 / INSTAGRAM

Four columns expose the real difference.

Watch & read
LinkedIn
DAY 05 / LINKEDIN

Agentic speed needs a proof system.

Watch & read
TikTok
DAY 05 / TIKTOK

The fastest demo can hide the wrong authority.

Watch & read
X
DAY 05 / X

A hook list is not a control map.

Watch & read
Instagram
DAY 06 / INSTAGRAM

Green needs a provenance trail.

Watch & read
LinkedIn
DAY 06 / LINKEDIN

The worker's tests are useful but interested.

Watch & read
TikTok
DAY 06 / TIKTOK

The terminal says zero. The suite did not pass.

Watch & read
X
DAY 06 / X

Exit zero is not enough evidence.

Watch & read
Instagram
DAY 07 / INSTAGRAM

More reviewers may mean one repeated premise.

Watch & read
LinkedIn
DAY 07 / LINKEDIN

An agent chat is not a decision record.

Watch & read
TikTok
DAY 07 / TIKTOK

A reviewer can inherit the author's favorite proof.

Watch & read
X
DAY 07 / X

Two approvals are not two observations.

Watch & read
Instagram
DAY 08 / INSTAGRAM

The polished review hides blank cells.

Watch & read
LinkedIn
DAY 08 / LINKEDIN

A model loop won't make you an engineer.

Watch & read
TikTok
DAY 08 / TIKTOK

Both models approved the backfill.

Watch & read
X
DAY 08 / X

Two models are not two sensors.

Watch & read
Instagram
DAY 09 / INSTAGRAM

The green UI has five blank cells.

Watch & read
LinkedIn
DAY 09 / LINKEDIN

Can the builder explain the failure path?

Watch & read
TikTok
DAY 09 / TIKTOK

The backfill passed review. Restart it.

Watch & read
X
DAY 09 / X

A demo can pass with no state model.

Watch & read
Instagram
DAY 10 / INSTAGRAM

What survives a context reset?

Watch & read
LinkedIn
DAY 10 / LINKEDIN

Durable notes are necessary, not ownership.

Watch & read
TikTok
DAY 10 / TIKTOK

The agent that built it is gone.

Read the full note
X
DAY 10 / X

A summary is not a causal model.

Watch & read
Instagram
DAY 11 / INSTAGRAM

The cost curve starts after launch.

Watch & read
LinkedIn
DAY 11 / LINKEDIN

AI makes building cheap, not owning.

Read the full note
TikTok
DAY 11 / TIKTOK

The founder built the tool in a weekend.

Watch & read
X
DAY 11 / X

The demo is the cheapest part.

Watch & read
Instagram
DAY 12 / INSTAGRAM

Three designs need three predictions.

Watch & read
LinkedIn
DAY 12 / LINKEDIN

Let models broaden the option set.

Read the full note
TikTok
DAY 12 / TIKTOK

The old incident changes the answer.

Watch & read
X
DAY 12 / X

Two models can widen the options.

Watch & read
Instagram
DAY 13 / INSTAGRAM

Map each assumption to one fault.

Watch & read
LinkedIn
DAY 13 / LINKEDIN

Record the prediction before the result.

Watch & read
TikTok
DAY 13 / TIKTOK

The canary makes both models revise.

Watch & read
X
DAY 13 / X

What would prove the design wrong?

Read the full note
Instagram
DAY 14 / INSTAGRAM

The missing option can be the costly one.

Watch & read
LinkedIn
DAY 14 / LINKEDIN

Confidence is routing input, not permission.

Watch & read
TikTok
DAY 14 / TIKTOK

Rename the labels. Does the policy flip?

Watch & read
X
DAY 14 / X

Type-safe is not decision-safe.

Watch & read
Instagram
DAY 15 / INSTAGRAM

A local model can win where the network loses.

Read the full note
LinkedIn
DAY 15 / LINKEDIN

Start routing with the job.

Read the full note
TikTok
DAY 15 / TIKTOK

The Jev-plus-LLM tax is easy to hide.

Read the full note
X
DAY 15 / X

The fastest call can make a slower answer.

Read the full note
Instagram
DAY 16 / INSTAGRAM

Map the control to the layer it owns.

Read the full note
LinkedIn
DAY 16 / LINKEDIN

Build a certainty ladder for agent work.

Read the full note
TikTok
DAY 16 / TIKTOK

A prompt cannot revoke a production key.

Read the full note
X
DAY 16 / X

A grammar can force shape, not truth.

Read the full note
Instagram
DAY 17 / INSTAGRAM

A router should know when to stay out.

Read the full note
LinkedIn
DAY 17 / LINKEDIN

Decision logs do not make an agent self-improving.

Read the full note
TikTok
DAY 17 / TIKTOK

A valid choice can go stale before use.

Read the full note
X
DAY 17 / X

Whose loop is your hook actually in?

Read the full note
Instagram
DAY 18 / INSTAGRAM

The agent chart can go up while delivery stalls.

Read the full note
LinkedIn
DAY 18 / LINKEDIN

Buy owner bandwidth before agent seats.

Read the full note
TikTok
DAY 18 / TIKTOK

The team doubled workers. Decisions got later.

Read the full note
X
DAY 18 / X

More agents can mean less shipped work.

Read the full note
Instagram
DAY 19 / INSTAGRAM

One evidence card can beat a giant prompt.

Read the full note
LinkedIn
DAY 19 / LINKEDIN

Deliver context when decisions need it.

Read the full note
TikTok
DAY 19 / TIKTOK

Kill the agent session halfway.

Read the full note
X
DAY 19 / X

Put evidence in the handoff.

Read the full note
Instagram
DAY 20 / INSTAGRAM

Engineers still own five decisions.

Read the full note
LinkedIn
DAY 20 / LINKEDIN

As code volume rises, judgment gets scarcer.

Read the full note
TikTok
DAY 20 / TIKTOK

Who owns the generated patch?

Read the full note
X
DAY 20 / X

The engineer becomes the decision owner.

Read the full note
Instagram
DAY 21 / INSTAGRAM

Kimi K3 is not a Mac mini workload.

Read the full note
LinkedIn
DAY 21 / LINKEDIN

Full-model service is a fabric problem.

Read the full note
TikTok
DAY 21 / TIKTOK

Kimi K3 on a Mac mini? Show me the bytes.

Read the full note
X
DAY 21 / X

Kimi K3 is not going on your Mac mini.

Read the full note