Notes
Short pieces about the methodology and architecture decisions behind the AI systems I ship — specs, evals, multi-agent orchestration, LLM integration, and the discipline of directing coding agents.
July 3, 2026
The agent that rewrites itself has no spec
The new pitch is the 'self-evolving' agent — it adapts its own behavior as it runs, so you never have to update it. ServiceNow and NVIDIA just aimed one at your desktop. But production software needs one thing above all else: a fixed target you can verify against. An agent that rewrites its own logic is a system where the thing you tested is not the thing running next week. Let it learn in data you can read and reset — never in behavior you can't watch change.
- agents
- architecture
July 1, 2026
Double the task, quadruple the failure
Everyone wants the agent that works a full 8-hour day. The math is against it. A new 2026 paper shows that doubling a task's length doesn't double the failure rate — it roughly quadruples it, because a tiny per-step error compounds. A 2% slip per step becomes a 33% chance of blowing the whole task over 20 steps. Long-horizon autonomy isn't waiting for a smarter model. It's an architecture problem: decompose, checkpoint, verify.
- agents
- architecture
July 1, 2026
Per-seat pricing is dead. Most founders are mispricing the corpse.
Agents do work without sitting in a seat, so per-seat SaaS is economically broken — seat-based pricing is already sliding and outcome-based pricing is the hot replacement. But the reflexive jump to 'charge per outcome' is a trap for anyone whose agent isn't reliable enough to guarantee the outcome. The real principle isn't a pricing fad; it's that your pricing model should follow who bears the reliability risk. It's a confession about how much you trust your own product.
- business
- agents
July 1, 2026
Perplexity is walking away from MCP — and they're not wrong
MCP won the standards war so fast that almost nobody stopped to ask whether it's actually good in production. Then Perplexity's CTO said out loud they're moving off it internally — because tool metadata can eat 40–50% of your context window before the agent does a single useful thing. The 'just plug in 50 MCP servers' dream collides with context economics. Tools are a dependency, not a buffet.
- architecture
- agents
July 1, 2026
The best agent is the most boring one
The industry sells autonomy as the goal: hand the agent a fuzzy objective and let it figure out the how. But the systems that actually survive production do the opposite — constrained tools, deterministic workflows, bounded decisions, human gates. Autonomy isn't a virtue you maximize; it's a budget you spend down, and every dollar you spend buys you a new way to fail. Put the intelligence in the narrow decision, and make everything around it dumb, predictable scaffolding.
- agents
- architecture
July 1, 2026
Voice agents just crossed the latency line
For years, AI voice agents failed on one thing: the pause. That half-second of dead air after you stopped talking made every phone bot feel broken. In 2026 the pause is gone — streaming speech end to end, new state-space voice models at 40ms, and sub-500ms round trips put voice inside the window where a conversation feels real. The model was never the hard part. Timing was — and timing is now an engineering problem, not a research one.
- ai-native
- agents