← Blog
Retrospective collection

Field notes

99 weekly reconstructions. “Week of” is the historical context; “Published” records the publication date. These were not published contemporaneously. Each entry distinguishes sourced developments from retrospective analysis.

2026

Week of
Link the proven parts

Build the missing piece after checking what already works.

Retrospective · Published
Week of
Be creative. Keep the essence.

The earlier two-hand logo brief remains a constraint on new image tools.

Retrospective · Published
Week of
Give me the diff and the checks

Coding work should arrive ready to review.

Retrospective · Published
Week of
Search for the skills

The October correction was simple: find the existing work first.

Retrospective · Published
Week of
Keep the old tests in the new evaluation

A generation change should not erase the tasks the application already depends on.

Retrospective · Published
Week of
Leave the missing field empty

Lightweight extraction should not invent completeness.

Retrospective · Published
Week of
Watch what people keep deleting

Repeated edits are evidence about the default model.

Retrospective · Published
Week of
Flat and lifeless is still flat and lifeless

The personal site needed a story and better assets, not more technical copy.

Retrospective · Published
Week of
Stop talking and take the correction

A live assistant needs to remain steerable while work runs.

Retrospective · Published
Week of
Turn one threat into a test

A report should lead to a specific check in the application.

Retrospective · Published
Week of
What owns the desktop agent?

The Hermes investigation carries forward a question about runtime responsibilities.

Retrospective · Published
Week of
What is worth keeping from a work session?

The capture idea starts with selecting moments, not accumulating footage.

Retrospective · Published
Week of
Look underneath HARPA

The research brief was about mechanisms worth adapting.

Retrospective · Published
Week of
The release calendar is not the roadmap

Test a new model when there is a reason to change the working system.

Retrospective · Published
Week of
Test it with the tools attached

A model evaluation does not cover every application built around it.

Retrospective · Published
Week of
The scene moved

Physical actions need current observations and a reliable stop.

Retrospective · Published
Week of
Leave a checkpoint someone can open

Long-running work needs saved results between the start and the final report.

Retrospective · Published
Week of
Find what we decided

Returning five old chats can leave the original question unanswered.

Retrospective · Published
Week of
Route by the cost of being wrong

Model selection should account for consequence as well as request size.

Retrospective · Published
Week of
Give it the outcome, then let it work

A clear brief should remove questions rather than generate more of them.

Retrospective · Published
Week of
We are already in a project

The Signal brief needed working files, not another explanation of where to build them.

Retrospective · Published
Week of
Put the address in writing

Voice answers need somewhere to leave the details.

Retrospective · Published
Week of
Keep the suspension in the timeline

A launch announcement alone does not describe availability for the whole week.

Retrospective · Published
Week of
Test the replacement before the deadline

A retired endpoint can expose dependencies hidden in old scripts.

Retrospective · Published
Week of
Finish the existing job first

A stronger model can remove friction without expanding the brief.

Retrospective · Published
Week of
Ask at the consequential part

Routine choices should not become a stream of approval requests.

Retrospective · Published
Week of
The laptop still has to be awake

Remote access needs to say where the work is actually running.

Retrospective · Published
Week of
What happens on the second webhook?

Deduplicate events before they produce duplicate actions.

Retrospective · Published
Week of
The preview became somebody’s daily tool

Dependency records should include the experiments people now rely on.

Retrospective · Published
Week of
Where is the actual file?

Completion should be visible in the deliverable.

Retrospective · Published
Week of
Keep the old route reachable

Know what reverting a model upgrade will actually restore.

Retrospective · Published
Week of
Show the finding and the fix

A security report should leave a maintainer with something they can verify.

Retrospective · Published
Week of
Who keeps it running?

Choosing a deployment also means choosing its maintenance work.

Retrospective · Published
Week of
The dispatcher does not need to be a town hall

clawREFORM needed clear technical names underneath its city analogy.

Retrospective · Published
Week of
Give the small model a small job

Bound the task and include the cost of escalation.

Retrospective · Published
Week of
The shot looks good. Does it cut?

Generated footage still has to work inside an edit.

Retrospective · Published
Week of
Which document is actually current?

A larger context window can hold more contradictory material.

Retrospective · Published
Week of
Reconnect to the work

A restored socket should recover the task, not just restart the animation.

Retrospective · Published
Week of
Stop using the flagship for string formatting

Split the workload before deciding which model should handle it.

Retrospective · Published
Week of
Can another worker finish this?

Save enough state that an interrupted task can be resumed without guesswork.

Retrospective · Published
Week of
Five agents, one unfinished task

Parallel work needs ownership at the point where the pieces come together.

Retrospective · Published
Week of
Remember the correction too

Personal context needs dates, scope and a way to fix it.

Retrospective · Published
Week of
What did latest change?

A moving model alias can change the product without a code deployment.

Retrospective · Published
Week of
Try the second provider

Compatible request shapes only get you halfway to portability.

Retrospective · Published
Week of
Change the background. Keep the product.

Image edits need a check on everything that was meant to stay untouched.

Retrospective · Published
Week of
Keep the difficult examples

Start with work that exposes a failure, then compare the models.

Retrospective · Published

2025

Week of
Keep the decisions. Drop the origin myth.

The year's announcements are easy to list. The project history is less tidy, which makes it more interesting.

Retrospective · Published
Week of
Open the thing you just made

Markdown isn't the PDF. Source code isn't the page. The exported result is where the person encounters the work.

Retrospective · Published
Week of
Nice formatting can hide a bad answer

GPT-5.2 was announced on 11 December, with professional work and artifacts prominent in the release. It's easy to see why polished spreadsheets, slides and documents get attention. They look finished.

Retrospective · Published
Week of
Send the result, not a tour of the session

A task can take hours and still deserve a short handoff.

Retrospective · Published
Week of
How do these projects actually fit together?

On 28 November, I was trying to restructure how AEGNTIC, AE Limited, Prompt Fail and the other projects worked together, then make that understandable on a website.

Retrospective · Published
Week of
Keep the project usable outside the agent IDE

Gemini 3 and Antigravity arrived on 18 November; Opus 4.5 followed on 24 November. More of the development experience was being organised around agents.

Retrospective · Published
Week of
A quiet week is allowed

There isn't a new release behind this entry. The source carried forward here is October's Haiku 4.5 announcement.

Retrospective · Published
Week of
The test has to be capable of failing

A file existing is a very low bar for a successful task.

Retrospective · Published
Week of
Don't make the fallback start from nothing

Model routing creates a handoff problem. When the first model can't finish, what exactly does the next one receive?

Retrospective · Published
Week of
Cheap calls. Expensive afternoon.

The API bill doesn't include the time spent fixing the output.

Retrospective · Published
Week of
Give the small model a complete small job

Haiku 4.5 arrived on 15 October. That makes model routing interesting, but I wouldn't start by replacing every expensive call with a cheaper one.

Retrospective · Published
Week of
Six agents editing one header is not a strategy

Parallelism is useful when the work separates cleanly. It gets messy when the split is just “give another agent something to do”.

Retrospective · Published
Week of
The handoff should fit on one screen

A transcript can contain everything and still be a terrible place to resume a task.

Retrospective · Published
Week of
Use the SDK. Keep responsibility for the job.

Anthropic introduced Sonnet 4.5 and the Claude Agent SDK on 29 September. Reusing maintained infrastructure is much more appealing than building another tool loop from scratch.

Retrospective · Published
Week of
The blog generator needs a reason to write

Automatic blog creation was part of the directory brief. I don't want that to mean a machine producing articles because Tuesday arrived.

Retrospective · Published
Week of
A price without its conditions is barely a price

An AI directory that lists “cheap” next to a tool hasn't done the comparison yet.

Retrospective · Published
Week of
No, that is still the same announcement

The LLM.deals brief called for automatic news and blog updates. That raises an immediate problem: ten sites can repeat one announcement and look like ten stories to a careless pipeline.

Retrospective · Published
Week of
LLM.deals needs to save someone a decision

On 1 September, I asked for LLM.deals as a directory: Angular, automatic updates, a news feed and blog creation. That was the brief. It wasn't a claim that all of it already worked.

Retrospective · Published
Week of
Same request shape. Different behaviour.

DeepSeek-V3.1 arrived on 21 August. A familiar API makes trying another model much easier. It doesn't make migration a one-line engineering decision.

Retrospective · Published
Week of
Put the new model in the old project

GPT-5 was announced on 7 August. The obvious thing to do next is try it. The less obvious thing is to resist making the test unusually easy.

Retrospective · Published
Week of
Can it run here, and can it actually do anything?

OpenAI released gpt-oss on 5 August and GPT-5 on 7 August. On 9 August, my questions were about local Ollama models that could work with terminal tools, then about the new gpt-oss models.

Retrospective · Published
Week of
Actually, use the other one

Long tasks get corrected halfway through. That's normal. Someone remembers a constraint, changes an account, or decides the first option wasn't right after all.

Retrospective · Published
Week of
The page is not your boss

ChatGPT agent arrived on 17 July, bringing research and browser actions into one workflow. That puts a lot of ordinary-looking buttons between a request and its consequences.

Retrospective · Published
Week of
The portfolio should survive the app that inspired it

On 19 July, I brought in a VisualMind map of the AEGNTIC GitHub profile and asked to turn it into a portfolio. Then I added the part that mattered: it shouldn't depend on that app continuing to exist.

Retrospective · Published
Week of
Give it a shell. Give it a job.

Gemini CLI's June release raises a fairly ordinary question with fairly large consequences: what exactly did we give the agent permission to do?

Retrospective · Published
Week of
The build passed. The header is still covering the article.

Gemini CLI landed on 25 June. AI in the terminal makes sense to me: the files, commands and tests are already there. Moving between a chat window and the actual project is mostly friction.

Retrospective · Published
Week of
Make the project portable between assistants

Gemini CLI adds another reason to keep instructions and checks in the repository.

Retrospective · Published
Week of
Build the path before the platform

Add infrastructure when a real task needs it.

Retrospective · Published
Week of
Which test passed?

A green result needs a named behaviour behind it.

Retrospective · Published
Week of
Explain it without making people feel behind

The June AEGNTIC brief wanted technical depth that newcomers could follow.

Retrospective · Published
Week of
Show every example

The design brief wanted visible effects, reduced-motion support and a real performance check.

Retrospective · Published
Week of
Leave enough information to resume

Longer assignments need checkpoints that describe actual state.

Retrospective · Published
Week of
What would make this coding task finished?

The Codex preview makes the acceptance criteria worth writing down.

Retrospective · Published
Week of
Accept the idea. Set up the project.

The May workflow request connected research, documentation, project setup and a shared app identity.

Retrospective · Published
Week of
Have the checkout ready

The video workflow brief asked the agent to handle preparation and make the remaining handoff small.

Retrospective · Published
Week of
Research, tools and a reason to come back

The April site concept joined products, writing, email updates and a three-dimensional visual direction.

Retrospective · Published
Week of
The plan has to change when the evidence does

o3 and o4-mini arrived on 16 April. Tool results still need to steer the next action.

Retrospective · Published
Week of
Start with the sitemap

April planning asked how the AEGNTIC site should fit together.

Retrospective · Published
Week of
Shrink the preview before approving it

The full-size generation window is a flattering place to judge a small asset.

Retrospective · Published
Week of
Change this bit. Keep the rest.

Image editing needs an explicit list of what stays unchanged.

Retrospective · Published
Week of
What does success mean in this tool?

A timeout and a rejected request need different recovery paths.

Retrospective · Published
Week of
Connect a few tools properly

The Responses API, Agents SDK and Gemma 3 arrived in the same week.

Retrospective · Published
Week of
Some tasks should just be quick

Spend extra reasoning on ambiguity and costly mistakes.

Retrospective · Published
Week of
Give the coding agent a usable project

Claude Code arrived alongside another week of model releases. The repository still needs to explain itself.

Retrospective · Published
Week of
Enough research. What are we trying next?

A research brief should make the next experiment easier to choose.

Retrospective · Published
Week of
Stop asking me about every click

Browser agents need a clear scope and a sensible point to interrupt.

Retrospective · Published
Week of
A link should take you to the evidence

Source links need to support the sentence beside them.

Retrospective · Published
Week of
Where did that answer come from?

Deep research made multi-step investigation more visible. Its claims still need checking.

Retrospective · Published
Week of
The browser has the final say

R1 and Operator put reasoning and execution beside each other.

Retrospective · Published
Week of
Run the awkward version of the demo

A clean example shows possibility. Messy inputs show where the work remains.

Retrospective · Published
Week of
Looks real. Does it behave properly?

NVIDIA Cosmos puts the difference between a convincing scene and a dependable system in view.

Retrospective · Published
Week of
Cheap tokens, expensive interruptions

Compare the cost of a finished task, including the repairs.

Retrospective · Published

2024