Welcome to the first issue of Agentic Coding Weekly, your filter for AI-assisted coding tools and agentic workflows without the AI hype.
Here’s what mattered in agentic coding this week 🙌 .
Latest Developments
Gemini 3: Google released their SOTA Gemini 3 model. It tops almost all the LLM benchmarks including the LMArena WebDev leaderboard, LiveCodeBench Pro, and Terminal Bench 2.0. It performs closely to SOTA on SWE-bench verified.
Pricing at $2/$12 per million tokens sits between Gemini 2.5 Pro ($1.25/$10) and Claude Sonnet 4.5 ($3/$15).
Default high thinking level enabled. Knowledge cutoff: January 2025.
Gemini CLI now supports Gemini 3.
Antigravity: Along with Gemini 3, Google launched Antigravity, their AI powered IDE.
Not mentioned anywhere in branding but it's clearly a VS Code fork.
Demo showed a pretty powerful example workflow loop of creating mockups with Nano Banana Pro, having agents implement the code, and verifying via browser use. Users can iterate by adding comments directly to screenshots. If they nail this kind of workflow, I might be convinced to use to Antigravity over Claude Code for prototyping.
I gave it a try for couple of prompts. Definitely rough around the edges for now. Users also report running out of credits within a few prompts with no current payment method to extend usage.
GPT-5.1-Codex-Max: A day after Gemini 3 release, OpenAI released their SOTA coding model. Beats Gemini 3 Pro and Claude Sonnet 4.5 on SWE bench verified and Terminal bench 2.0.
Supposed to be more token efficient. Uses 30% fewer thinking tokens than GPT-5.1-Codex at medium reasoning effort while achieving better performance.
Built for long-running, detailed work. Uses "compaction" to operate across multiple context windows, handling millions of tokens for a single task.
Available with ChatGPT Plus, Pro, Business, Edu, and Enterprise plans. Not yet in API.
Playbook of the Week
Use smartcommit to generate git commit message. It's a middle ground between writing detailed commit message yourself and letting AI completely generate commit messages based on the diff without any of your input.
smartcommit analyzes staged changes, asks clarifying questions about your code changes, and generates structured commit messages.
Articles and Ideas
On The Pragmatic Engineer podcast, Martin Fowler discussed the fundamental shift software engineering is undergoing, moving from deterministic systems (testing, refactoring) to non-deterministic LLM-based workflows.
The section about vibe coding caught my attention. Sounds quite obvious but it does need saying it out loud: Vibe Coding is detrimental to long-term software quality because it eliminates the crucial learning loop for developers. This is the (almost) exact quote from the episode:
If you’re not looking at the output, you’re not learning. The thing is that so much of what we do is we come up with ideas, we try them out on the computer with this constant back and forth between what the computer does with what we’re thinking. We’re constantly going through that learning loop. And Unmesh’s point, which I think is absolutely true, is you cannot shortcut that process. What LLMs do is they just kind of skim over all of that, and you’re not learning. And when you’re not learning, that means that when you produce something, you don’t know how to tweak it, modify it, evolve it, and grow it.
The AI coding market is splitting into two segments:
"Hands-off" tools for non-engineers focused on prototyping like Lovable
"Hands-on" tools for professional engineers using AI as a pair programmer like Claude Code
Key takeaway seems to be the analogy of AI with early compilers:
Just as compilers freed programmers from writing assembly code, AI tools are freeing software engineers from the grunt work of writing boilerplate and routine code, and letting them us focus on higher-order thinking.
An argument against monolithic MCP servers. The alternative: minimal CLI tools executed via Bash.
For his use case, author used minimal Node.js scripts wrapped in Bash commands. Agents already understand Bash. This approach reduced token consumption from 18k to 225 tokens in his tests.
Recommendation: only add essential MCP servers to your project. Better yet, implement needed tools yourself to keep context clean.
That's it for this week. Next issue drops Monday.