Matt Stark

[Matt Stark @thastark]

@trq212 I hit 28% weekly Fable usage maxing session limit an hour and a half before reset with 20% weekly usage in that tiny reset window. My harness delegates to sub-agents: seems high IMO. New repo so no evals and Fable likely took on more but orchestrator role is usually an 8% job.

[Matt Stark @thastark]

@waldekm Is a custom harness for Claude/Copilot/Codex. Can specify model using frontmatter in your md's. I find what works then encapsulate in a custom slash command (you call skills) to build repeatable procedures within each repo. The harness is an agentic wrapper to enforce repo rules.

[Matt Stark @thastark]

@waldekm With this being said, IMO our future is finding the most token efficient least-most capable model for the task. Thats an output judgement. Economics will rule and model choice is important. I am OK using Haiku to update my wiki but (today) wouldn't use it to improve performance.

[Matt Stark @thastark]

@waldekm My harness still delegates to the least-most capable model for the task to max token efficiency.

Opus 4.6->Opus 4.7 and Opus 4.8->Fable where orchestrator behaviour changes were big had the biggest impact on my outputs.

[Matt Stark @thastark]

@waldekm Opus 4.8->Fable requires less specificity in agentic instructions and TBH my harness seems like an impediment. Honesty contracts are less important but enforcing guardrails = same (now 5 sub agent layers supported). Orchestrator should DIY more. Some sub agents totally replaced.

[Matt Stark @thastark]

@waldekm Opus 4.6->4.7 required greater specificity in agentic instructions+wiki, honesty contracts between spec & code, and hooks to deterministically keep outputs between guardrails (only 2 sub agent layers supported). All orchestrator responsibility. Sub-agents keep using model B.

[Matt Stark @thastark]

@waldekm And owning vs not owning the API. Also where the tests are executed. Most days I dread looking at my GitHub Actions budget.

[Matt Stark @thastark]

@waldekm * Testing against 4** - is actually every possible http status code, on every possible surface.

[Matt Stark @thastark]

@waldekm Tests running against all 4** codes is another rule, and each must fail gracefully. Setup is more complex using canonical datasets, realistic API workflows / human journeys. DevProxy|middleware is the way I keep mocks out of code. Backoff is easy, being comprehensive is not.

[Matt Stark @thastark]

@waldekm I've found Claude to be comprehensively honest ~15% of the time, ~10% of my implementation loop is actually building, the other ~90% is strict guardrails (time measurement). I am well over 4k deterministic tests on a tool with ABAC ~20 roles/100s of perms (jest & playwright).

[Matt Stark @thastark]

@RoyKimYYZ Congrats Roy!

[Matt Stark @thastark]

@newsystems_ brampton

[Rowan Cheung @rowancheung]

JUST IN: Sam Altman's firing was driven by OpenAI's secret breakthrough AI named Q* (possibly Q-learning).

Plus, huge developments in AI today from Inflection AI, Google Bard, Neuralink, ElevenLabs, screenshot-to-code, and 9 new AI tools.

Here's EVERYTHING you need to know:

[Matt Stark @thastark]

@GoPro do you support streaming multiple cameras? We are a hockey team, and would use many cameras to capture games. Cameras like @Pix4Team and @veocamera are great, but don’t capture details close to the net. Do you have options or even third party software you could recommend?

[Microsoft 365 Developer @Microsoft365Dev]

Learn how to build and sell applications for Microsoft Teams at the #MicrosoftTeamsAppCamp:

https://t.co/mAYJWGTGLm

#Microsoft365Dev #MicrosoftTeams

[Matt Stark @thastark]

@Bell is pathetic. Installs wrong TV package then bills me the activation fee said to be waived with online orders. Their WiFi router nukes cell reception in mine and each neighbour’s home. When I complain they try to increase my bill by $70. @CRTCeng are you protecting us? #help

[Matt Stark @thastark]

@cmcnulty2000 @RoyKimYYZ The reality is, with all the open source cloud libraries, it makes a lot of sense to take a scripted approach. Mover seriously lacks APIs and hence impedes any realistic / reasonable change management process. We’ve already got converters open source. No value add here!

[Matt Stark @thastark]

@cmcnulty2000 @RoyKimYYZ Creating a new UI really doesn’t do much to help migrate data. I’d like to see bilateral sync, and shared folder connectors. Anyone who’s used Mover knows it’s difficult if not impossible to do everything at once. We are using mover then Power Automate to reverse the sync...

[Matt Stark @thastark]

@RoyKimYYZ #m365