Programming & AI Tips πŸ’‘
45.8K subscribers
67 photos
11 videos
30 files
388 links
Programming & AI:
Tips πŸ’‘
Articles πŸ“•
AI & LLMs πŸ‘Ύ
Design, Architecture, Principles βœ…

πŸ‡³πŸ‡± Contact: @MoienTajik
Download Telegram
Make coding agents earn the green check πŸ’‘

Do not end an agent task with "implement this." End it with a command that can fail.

For a .NET repo, that might be dotnet test, a focused test project, a formatter check, or a small reproduction script. Tell the agent which command to run, what output to inspect, and what it should do if the command fails.

A good task contract has three parts:

β€’ Change: the files or behavior to update.
β€’ Proof: the exact command or test case.
β€’ Stop condition: what needs human review instead of another retry.

This also makes PR review faster. You get a diff plus evidence, not a confident paragraph saying the fix should work.

γ€°γ€°γ€°γ€°γ€°γ€°
#AI #dotnet #Testing
@ProgrammingTip
GitHub moved the Copilot runtime to Rust πŸš€

GitHub migrated the runtime behind Copilot to Rust, and used Copilot during the migration itself.

βœ… Why this is worth reading:

β€’ Runtime work is systems work: a language migration means tracing real production behavior, not just translating syntax.

β€’ Copilot was part of the workflow: useful for exploring an unfamiliar codebase, drafting changes, and keeping momentum through repetitive conversion work.

β€’ The output still needs engineering judgment: boundaries, performance, correctness, and rollout safety are not autocomplete problems.


This is a practical case study for teams asking where coding agents help on a large refactor. Use them to accelerate investigation and implementation. Keep humans on architecture, tests, and production checks.


[ Blog ] :
https://github.blog/ai-and-ml/generative-ai/migrating-the-github-copilot-runtime-to-rust-using-copilot

〰️〰️〰️〰️〰️〰️
#AI #GitHub #Rust #Copilot
@ProgrammingTip
Please open Telegram to view this post
VIEW IN TELEGRAM
Gemini 3.8 Live is generally available πŸš€

Google moved Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking to general availability in the Gemini API.

These are the real-time audio models exposed through the Live API. Live is for the voice path: streaming audio in, streaming audio out, and handling a conversation without bolting together STT, an LLM call, and TTS yourself.

What to check:

β€’ Gemini 3.8 Live: the lower-latency voice model.
β€’ Extended Thinking: use it when the spoken request needs more reasoning before the reply.
β€’ GA status: worth revisiting if you held off on a preview-only voice feature.

For .NET teams, this is a good fit for a streaming WebSocket service, not a request-response controller.

[ Read More ] :
https://ai.google.dev/gemini-api/docs/changelog

γ€°γ€°γ€°γ€°γ€°γ€°
#AI #Gemini #LLM #API
@ProgrammingTip
Hex turns GPT-6 Astra analysis into visual reports πŸš€

Hex is using GPT-6 Astra to turn complex analysis into visual reports. The useful bit is not just asking a model to summarize a table. It is moving from an analysis request to a result people can inspect and share.

What this points to:

β€’ Analysis as an artifact: teams need charts, assumptions, and outputs, not a chat answer pasted into Slack.
β€’ Human review still matters: a clean report can hide bad joins, stale data, or a wrong metric definition.
β€’ Tool context is the product: models get more useful when they operate inside the workspace where data and business logic already live.

For AI app builders, this is the bar: produce a result that can survive review, not just a plausible paragraph.

[ Read More ] :
https://openai.com/index/hex-gpt-6-astra

γ€°γ€°γ€°γ€°γ€°γ€°
#AI #LLM #GPT6 #Data
@ProgrammingTip
GPT-6 Sol and GPT-6 Luna are in the API πŸš€

OpenAI shipped GPT-6 Sol and GPT-6 Luna for API developers, alongside their availability in Codex and ChatGPT.

This is a two-model release, so do not blindly swap your existing production model. Put both behind the same eval set first: tool calls, structured output, long-context retrieval, refusal behavior, and latency under your real prompt size.

What to do this week:

β€’ Add Sol and Luna as versioned model options in your config.
β€’ Run replay traffic against a fixed golden set.
β€’ Log model ID, token use, tool errors, and task success separately.

A model migration is an engineering change, not a dropdown change.

[ Read More ] :
https://openai.com/index/introducing-gpt-6-sol-and-luna/

γ€°γ€°γ€°γ€°γ€°γ€°
#AI #OpenAI #LLM #API
@ProgrammingTip
Claude Opus 5.5 is now on the Claude Platform πŸš€

Anthropic introduced Claude Opus 5.5, a lower-cost frontier model available in Claude Code and through the Claude Platform.

That matters if your agent workload has been split between a high-end model for hard tasks and cheaper models for everything else. A lower-cost Opus tier can change where that handoff happens.

What to check:

β€’ Run your existing eval set, not a few cherry-picked prompts.
β€’ Measure tool-call accuracy and recovery after a failed call.
β€’ Compare total agent cost: tokens, retries, and human review time.

For Claude Code users, model choice is now a practical repo-level config decision, not just a benchmark chart.

[ Blog ] :
https://www.anthropic.com/claude-opus-5-5

γ€°γ€°γ€°γ€°γ€°γ€°
#AI #LLM #Claude #ClaudeCode
@ProgrammingTip
Claude Opus 5.5 is available in the API πŸš€

Anthropic released Claude Opus 5.5, and it is available through the Claude API.

The practical pitch is simple: lower typical token costs than Opus 5, plus a faster mode when response time matters more than squeezing out the last bit of reasoning.

What to check in your evals:

β€’ Run the same tool-use and coding tasks against Opus 5.
β€’ Measure latency separately for normal and faster mode.
β€’ Track input and output tokens, not just the model's listed price.

A cheaper high-end model changes agent architecture decisions. Some workflows that needed routing to a smaller model may now fit under one stronger default.

[ Read More ] :
https://www.anthropic.com/claude/opus

γ€°γ€°γ€°γ€°γ€°γ€°
#AI #LLM #Claude #API
@ProgrammingTip
Claude Sonnet 5.5 is built for coding agents πŸš€

Anthropic released Claude Sonnet 5.5 for the Claude Platform. It targets the work developers actually hand to coding agents: navigating a repo, making changes across files, using tools, and checking the result.

What changed:

β€’ Faster agentic work: aimed at shorter tool loops and less idle time while an agent investigates a codebase.
β€’ Lower-cost option: positioned for teams that need to run coding tasks repeatedly, not just ask one-off questions.
β€’ Production focus: Anthropic calls out software engineering and multi-step agent workflows directly.

If your agent spends more time calling tools than writing code, model latency and per-task cost matter as much as benchmark scores. Test it on a real issue queue, with your actual tool permissions.

[ Read More ] :
https://www.anthropic.com/claude-sonnet-5-5

γ€°γ€°γ€°γ€°γ€°γ€°
#AI #LLM #Claude #CodingAgents
@ProgrammingTip
OpenAI introduced Dots, always-on agents πŸš€

OpenAI introduced Dots, its take on always-on agents.

This is not a one-prompt, one-answer workflow. The pitch is an agent that can stay active around work instead of waiting for you to reopen a chat and restate the task.

Why this matters:

β€’ Long-running work: agents need durable context, not a pile of copied prompts.
β€’ Real handoff points: a useful agent should surface decisions and results, not silently keep doing things.
β€’ Agent ops: permissions, logs, cancellation, and cost limits become product features.

If you build agent workflows, the hard part is no longer getting a model to call a tool. It is making an autonomous process observable enough that somebody will trust it.

[ Read More ] :
https://openai.com/index/introducing-dots/

γ€°γ€°γ€°γ€°γ€°γ€°
#AI #Agents #OpenAI
@ProgrammingTip
GPT-6 Astra gets an Ultrafast API tier πŸš€

OpenAI added an Ultrafast speed tier for GPT-6 Astra in Codex and the API.

The important bit is real-time work. Astra can now run Responses API workflows over WebSockets, which is a much better fit for interactive coding assistants, live agent status, and UI flows where waiting on a full request feels bad.

What to check:

β€’ Responses API: use it for the agent loop and tool calls.
β€’ WebSockets: keep one live connection instead of polling.
β€’ Ultrafast tier: test it where latency matters more than squeezing every last token of quality.

If your app streams agent work to a browser, this is worth benchmarking against your current model setup.

[ Article ] :
https://community.openai.com/t/build-ultrafast-with-astra-in-codex-and-the-api/1402393

γ€°γ€°γ€°γ€°γ€°γ€°
#AI #OpenAI #API #LLM
@ProgrammingTip