AI Conf 2026: Classic ML Is Dead, Everyone's Building Agents
Spent two days at AI Conf in Moscow. The shift is complete: nobody talks about traditional ML anymore. It's all agents, RAG, and voice systems.
1. Academic Publication Pipeline Is Slow
Average time from submission to publication at A-tier conferences: 9 months. Multiple review cycles, sequential improvements.
What researchers actually use LLMs for now:
- Code generation
- Paper review assistance
- Literature synthesis
- (Not for original ideas — tried "let it think for 2 weeks," expensive and ineffective)
Prediction: Future papers will include zip archives of experimental code that AI can verify. Human value shifts to idea generation, not implementation.
2. Search Agents Workshop
Built a working ReAct search agent in the workshop:
- Groq API — free tier, fast inference
- Tavily — 1000 free search queries/month
- Langfuse monitoring
Stack cost: $0 for prototyping. Production cost: depends on scale.
3. Monitoring: Langfuse vs Arize Phoenix
| Tool | Approach | Best For |
|---|---|---|
| Langfuse | Manual integration, detailed traces | Custom setups, granular control |
| Arize Phoenix | Auto-instrumentation, wraps everything | Quick setup, less configuration |
Both show traces, token counts, latency breakdowns. Phoenix wins if you want observability without wiring it yourself.
4. Agent Harness vs Classic Agents
The terminology evolved:
- 2024: "What's the difference between LLM and agent?"
- 2026: Agent Harness — memory + skills instead of tools
Example: Deep Agents framework. Skill creation costs 2M tokens. Single invocation: 100K tokens. But the abstraction is cleaner than manual tool orchestration.
5. Voice Agents for Telephony
Voice-to-voice models exist but lack:
- Tool use integration
- Context management
- Reliability for long conversations
Current production stack: Speech-to-Text → LLM → Text-to-Speech
Voice-to-voice will replace this eventually, but not before tool calling and context compression catch up.
What I Didn't Hear
- Gradient boosting use cases
- Feature engineering debates
- Model interpretability discussions (except for RAG context windows)
The industry moved on. If you're still pitching Random Forest improvements, you're talking to the wrong audience.
My Take
The conference confirmed what I see in production: agent orchestration is the new infrastructure layer. Not the models themselves — how you connect them, manage memory, route between skills, and monitor everything.
The companies winning aren't those with the best single model. They're those with the best agent architecture.
More production AI insights and conference notes from a bank's DS lead — follow my Telegram channel:
https://t.me/ai_tablet (Russian, technical)
More AI engineering notes, RAG benchmarks, and production insights from inside a bank — follow my Telegram channel:
🚀 https://t.me/ai_tablet (Russian, technical)