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
Build the missing piece after checking what already works.
The earlier two-hand logo brief remains a constraint on new image tools.
Coding work should arrive ready to review.
The October correction was simple: find the existing work first.
A generation change should not erase the tasks the application already depends on.
Lightweight extraction should not invent completeness.
Repeated edits are evidence about the default model.
The personal site needed a story and better assets, not more technical copy.
A live assistant needs to remain steerable while work runs.
A report should lead to a specific check in the application.
The Hermes investigation carries forward a question about runtime responsibilities.
The capture idea starts with selecting moments, not accumulating footage.
The research brief was about mechanisms worth adapting.
Test a new model when there is a reason to change the working system.
A model evaluation does not cover every application built around it.
Physical actions need current observations and a reliable stop.
Long-running work needs saved results between the start and the final report.
Returning five old chats can leave the original question unanswered.
Model selection should account for consequence as well as request size.
A clear brief should remove questions rather than generate more of them.
The Signal brief needed working files, not another explanation of where to build them.
Voice answers need somewhere to leave the details.
A launch announcement alone does not describe availability for the whole week.
A retired endpoint can expose dependencies hidden in old scripts.
A stronger model can remove friction without expanding the brief.
Routine choices should not become a stream of approval requests.
Remote access needs to say where the work is actually running.
Deduplicate events before they produce duplicate actions.
Dependency records should include the experiments people now rely on.
Completion should be visible in the deliverable.
Know what reverting a model upgrade will actually restore.
A security report should leave a maintainer with something they can verify.
Choosing a deployment also means choosing its maintenance work.
clawREFORM needed clear technical names underneath its city analogy.
Bound the task and include the cost of escalation.
Generated footage still has to work inside an edit.
A larger context window can hold more contradictory material.
A restored socket should recover the task, not just restart the animation.
Split the workload before deciding which model should handle it.
Save enough state that an interrupted task can be resumed without guesswork.
Parallel work needs ownership at the point where the pieces come together.
Personal context needs dates, scope and a way to fix it.
A moving model alias can change the product without a code deployment.
Compatible request shapes only get you halfway to portability.
Image edits need a check on everything that was meant to stay untouched.
Start with work that exposes a failure, then compare the models.
2025
The year's announcements are easy to list. The project history is less tidy, which makes it more interesting.
Markdown isn't the PDF. Source code isn't the page. The exported result is where the person encounters the work.
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.
A task can take hours and still deserve a short handoff.
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.
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.
There isn't a new release behind this entry. The source carried forward here is October's Haiku 4.5 announcement.
A file existing is a very low bar for a successful task.
Model routing creates a handoff problem. When the first model can't finish, what exactly does the next one receive?
The API bill doesn't include the time spent fixing the output.
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.
Parallelism is useful when the work separates cleanly. It gets messy when the split is just “give another agent something to do”.
A transcript can contain everything and still be a terrible place to resume a task.
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.
Automatic blog creation was part of the directory brief. I don't want that to mean a machine producing articles because Tuesday arrived.
An AI directory that lists “cheap” next to a tool hasn't done the comparison yet.
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.
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.
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.
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.
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.
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.
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.
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.
Gemini CLI's June release raises a fairly ordinary question with fairly large consequences: what exactly did we give the agent permission to do?
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.
Gemini CLI adds another reason to keep instructions and checks in the repository.
Add infrastructure when a real task needs it.
A green result needs a named behaviour behind it.
The June AEGNTIC brief wanted technical depth that newcomers could follow.
The design brief wanted visible effects, reduced-motion support and a real performance check.
Longer assignments need checkpoints that describe actual state.
The Codex preview makes the acceptance criteria worth writing down.
The May workflow request connected research, documentation, project setup and a shared app identity.
The video workflow brief asked the agent to handle preparation and make the remaining handoff small.
The April site concept joined products, writing, email updates and a three-dimensional visual direction.
o3 and o4-mini arrived on 16 April. Tool results still need to steer the next action.
April planning asked how the AEGNTIC site should fit together.
The full-size generation window is a flattering place to judge a small asset.
Image editing needs an explicit list of what stays unchanged.
A timeout and a rejected request need different recovery paths.
The Responses API, Agents SDK and Gemma 3 arrived in the same week.
Spend extra reasoning on ambiguity and costly mistakes.
Claude Code arrived alongside another week of model releases. The repository still needs to explain itself.
A research brief should make the next experiment easier to choose.
Browser agents need a clear scope and a sensible point to interrupt.
Source links need to support the sentence beside them.
Deep research made multi-step investigation more visible. Its claims still need checking.
R1 and Operator put reasoning and execution beside each other.
A clean example shows possibility. Messy inputs show where the work remains.
NVIDIA Cosmos puts the difference between a convincing scene and a dependable system in view.
Compare the cost of a finished task, including the repairs.