We took a contrarian bet on how to handle context. Instead of following the industry standard of using the primary model for compaction, we built a subagent system around Mercury 2. This led to 90% lower costs and 82% faster latency.
English
Ankur Rustagi
97 posts
At Augment, we aren’t tied to a single provider, which gives us the freedom to prioritize models with optimal speed and cost-efficiency for our users. Our recent experiments proved that Mercury 2 provides the ideal intelligence level for tasks like context compaction.
Inception@_inception_ai
@augmentcode rebuilt their context compaction layer around Mercury 2. 82% latency cut. 90% cost cut. Comparable quality to Opus 4.7. Running in production today. "We took a counter-intuitive bet. We decoupled summarization entirely, offloading it to Mercury 2 as a dedicated subagent. Mercury 2 is the highly efficient engine powering our most critical workflows." -@RustagiAnkur & @jm1234567890, Members of Technical Staff at Augment Code The subagent layer needs the most efficient model. Full methodology and eval setup in the writeup. inceptionlabs.ai/blog/rise-of-r…
English
@jitenoswal Saw the infinite loop issue on @firecrawl agent. Although seems like it's common with small context windows and low intelligence model. Good work.
English
Building AI agents with coding assistants and agent frameworks?
Your agent probably has:
— loops with no exit condition
— retries with no max limit
— tools your prompts reference but don't exist
— hardcoded API keys
— bad agent engineering patterns
No linter catches this. Code review misses it.
We built an open-source skill that does → Runs 100% locally.
Install with one command. Run inside your agent.
npx skills add aurite-ai/agent-verifier --all
Then say "verify agent" in your coding assistant.
OSS Github Repo: github.com/aurite-ai/agen…
Follow @jitenoswal for more OSS tools for building effective production Agents and for AI engineers.
GIF
English
Ankur Rustagi retweetledi
When you filter to PRs where someone actually engaged with the PR after the review, scores shift significantly.
@OpenAI’s Codex sees the largest swing at +17.2pp. @augmentcode climbs to #2 in F1 with +12.3pp.
Some tools score higher when nobody engages, others score higher when humans are actively involved. The delta reflects how each tool is being used, not which one is better.
English
@bayer04_en is a special place. When @kaihavertz29 scored the penalty, they didn't boo him. Infact celebrated him for doing so well in his football career. This is so wholesome to watch. Cheers Leverkusen!
Bayer 04 Leverkusen@bayer04_en
10 years at Bayer 04. Welcome back Kai. 🫶🖤❤️ #B04ARS
English
YouTube
English
Ankur Rustagi retweetledi
Inspired by @karpathy's brilliant microgpt.py — ported it to Rust. 208 lines, zero deps, ~400x faster
• Tape autograd (no graph/Rc)
• Flat Vec<f64>, cache-friendly
• No GC, just truncate
• Native math, no interpreter
• 283s → 0.68s
vinodbf16.substack.com/p/microgpt-in-…
English
@sama and @DarioAmodei can't shake hands. youtube.com/shorts/Z5aExYu…
Lessons:
- money and success does not guarantee peace and love.
- everyone thinks they are right. The right wrong arguments sometimes happen at world stages as well.
YouTube
English
@amruthagujjar Claude also tries to do the same. They need the money and recognition.
English
I do not like that cursor tries to add itself as a co-author on my PRs...
English
Everyone is projecting the first one person unicorn. Curious are there any one person unicorn/10 ($100m) startups sold yet?
English
Forgeon.dev beta is live . Best quality PDF generation that is just like your web dashboards. It's like @resend for creating a pdf. And it's 50-90% cheaper than any alternative. Enjoy!
English
@taheerBuilds @steipete @nutrientdocs AI can make it, I created forgeon.dev , highest quality PDF creation and 50-80% cheaper than any other player.
English
@steipete @nutrientdocs If you had to rebuild this with AI how much time would it take you??
English
I signed the contract today with software my company built years ago. (now it's @nutrientdocs, still best for anything PDF)
hari raghavan@haridigresses
Peter Steinberger bootstrapped PSPDFKit over 10 years (gold-standard PDF library used by Box, Apple, even DocuSign). Insanely hard tech. IYKYK. I'd guess he's worth a couple hundred mill from that. If he goes to OpenAI... he might make more from OpenClaw, which is 3 months old.
English
If you ever wondered whether agents talk to each other.
Two agents built at @augmentcode came up with there language ABX. And used it to identify, resolve and verify an issue. Today is just hackathon but integrated sdlc automation is the future.
English
Today is #hackathon day @augmentcode . We created an OpenClaw version internally and asked it to come up with an optimized language to talk to an older version agent. They came up with ABX (augment bot exchange) .
English
This is so cool! Check it out
Ankur Rustagi@RustagiAnkur
Shipped #Ollama support for MCPlexor 🚀 If your agent uses Linear + GitHub + Notion, you're dumping ~40k tokens into context on every request. That's 20% of a 200k context window gone to tools. MCPlexor fixes this. <1k tokens overhead. Dynamic routing. Try 100% local. Details 👇
English
Tomorrow, my colleagues at @augmentcode and I will be sharing about our SOTA code context engine. It improves @claudeai and @cursor_ai quality by ~70% .
Augment Code@augmentcode
Want to experience the magic of our Context Engine with your existing toolbase? Introducing Context Engine MCP: semantic indexing that works across Claude Code, Cursor, Zed, GitHub Copilot, and 10+ other agents. See an interactive live demo on Tuesday, February 10 at 10 AM PT. Implement in your workflows on the same day. Register now: watch.getcontrast.io/register/augme…
English
Who is MCPlexor for?
🔹 Power Users: Waitlist open for MCPlexor Cloud (managed routing, no infra).
🔹 Privacy / Local: Use the new Ollama backend. Zero cost, offline.
checkout: mcplexor.com
#AIAgents #MCP #Ollama #DevTools
English
Shipped #Ollama support for MCPlexor 🚀
If your agent uses Linear + GitHub + Notion, you're dumping ~40k tokens into context on every request. That's 20% of a 200k context window gone to tools.
MCPlexor fixes this. <1k tokens overhead. Dynamic routing. Try 100% local.
Details 👇
English
The solution: Semantic Multiplexing
MCPlexor sits between your agent and your tools.
Agent asks for "create issue"
We route to Linear
Only Linear tools load
Result: 95% token reduction.
And now, you can run the routing logic locally with Ollama (Llama 3, Mistral).
English