Cheaper tokens will not mean smaller AI budgets
OpenAI cut GPT-5.6 Luna prices by 80% and Terra by 20%. For big AI customers, lower unit costs are more likely to expand consumption than shrink the forecast.
// writing
Blog
Notes on agentic AI, multi-agent frameworks, and the occasional dispatch from the games that keep me up at night.
AI in games has to earn the player's suspension of disbelief
The industry is pitching living worlds while developers grow more hostile to generative AI. Both reactions make sense — they are talking about different products.
AI agents are becoming the new browser
If agents become the interface to the internet, tool access and recommendations become questions of competition — not just protocol design.
The flagship model is becoming a routing policy
GPT-5.6 arrived as three models and a multi-agent mode. The interesting trend is no longer one model winning — it is systems deciding how much intelligence a task deserves.
AI APIs are rented land
GitHub Models is shutting down. A useful reminder that the convenient layer in your AI stack is often the least permanent one.
What 2,000 hours of Dota taught me about building AI agents
Fog of war, item builds, and throwing from ahead — the best mental models I have for agentic systems came from a video game.
"Agent memory: beyond the context window"
Sessions, state, and long-term memory — how to make an agent that remembers without drowning in its own history.
Designing tools your agent won't misuse
Your agent is only as reliable as its worst tool. Field notes on tool ergonomics — the highest-leverage, least glamorous work in agentic AI.
"Evals for agents: stop shipping vibes"
You can't A/B test a feeling. How to build an evaluation harness for agentic systems before you touch another prompt.
Multi-agent patterns that survive contact with production
Orchestrator-workers, pipelines, and judges — the three multi-agent shapes I keep reusing, and the ones I've abandoned.
"Picking an agent framework in 2026: ADK, LangGraph, or no framework at all"
The agent framework wars, six months in — how I actually decide between Google ADK, LangGraph, the Claude Agent SDK, and a bare loop.
"MCP in practice: giving your agents hands"
The Model Context Protocol turned tool integration from N×M glue code into a plug standard. Notes from running it in anger.
"Google ADK first impressions: agents as code, finally"
A weekend spent porting a homegrown agent stack to Google's Agent Development Kit — what impressed me, what didn't.
Agents are just loops (and that's the hard part)
Strip away the hype and an AI agent is a while-loop around an LLM. The engineering is in everything the loop touches.