Devigner
Dev instincts. Designer eyes. No sacred rules.
Dev instincts.Designer eyes.No sacred rules.
When to use incremental tool updates instead of legacy streaming?
Codex tool updates on Google Cloud Vertex AI previously arrived in batches rather than as they happen; incremental sending is now the default, with a flag to opt out.
Read articleThe latest 16 of 78 articles, oldest to newest.
Articles
All articles- WorkflowCodex CLI
When to use incremental tool updates instead of legacy streaming?
Codex tool updates on Google Cloud Vertex AI previously arrived in batches rather than as they happen; incremental sending is now the default, with a flag to opt out.
- WorkflowVercel
When to use AI Gateway instead of direct image model APIs?
A separate MiniMax account, key, and bill just to generate product images slows you down until one gateway API key covers it.
- WorkflowVercel
How I filter agent tools without burning output tokens
Routing and policy checks with generative models spend tokens on text you never use.
- VerificationOpencode
When to use Copilot Auto instead of picking a model yourself?
Choosing a model for every session wastes time and money; Copilot Auto's router picks per session at a discount, with a fallback model if routing fails.
Thoughts
All thoughtsI let scripted opencode runs pile into my session list until finding the good conversations felt like inbox triage. One flag, --archived at launch, and machine runs stay out of sight by default. Tools that mix automation and human traffic need a separation knob from day one, not after the noise wins.
I keep blaming the model for slow agent loops, but the real bottleneck is my architecture making every step wait its turn. I want to rework my agents so steps run in parallel and cheap fast models handle the filler steps, saving the big model for the hard calls. Latency is a design choice, not a fact.
I lost a carefully written prompt once because I restarted my session at the wrong moment, and now I write long prompts in a scratch file before pasting them in. Tooling is finally catching up by preserving drafts across backend switches, but I won't trust that until I've tested it myself. The agent losing my work is annoying; me losing my own draft is worse.
Workflow field notes, straight to your inbox
Join readers getting workflow field notes and reusable assets in every issue. No spam, unsubscribe anytime.