Salamon Capital — Insights Research and analysis on AI productivity, measurement and automation. https://salamoncapital.com/insights 2026-08-20T00:00:00.000Z Salamon Capitalhq@salamoncapital.com The fastest iteration won. And still had to sell. https://salamoncapital.com/insights/cursor-and-the-fastest-iteration 2026-08-20T00:00:00.000Z Cursor changed what it was twice in under four years. It ended in a $60 billion acquisition by the company with the data centres. Both halves are the lesson. Price the agent before you build it https://salamoncapital.com/insights/agent-cost-calculator 2026-08-20T00:00:00.000Z The article on cost ends with advice to price one representative task. This is that advice as a tool. The harness, not the model https://salamoncapital.com/insights/the-harness-not-the-model 2026-08-17T00:00:00.000Z Firefox shipped 423 security fixes in a single month against a baseline of twenty. The reason is not the one most people take away from it. Subagents are not an org chart https://salamoncapital.com/insights/subagents-are-not-an-org-chart 2026-08-17T00:00:00.000Z Anthropic measured a 90% improvement from splitting one agent into several. The same guidance tells you to start with one — and names the split that ruins it. More agents is not more perspectives https://salamoncapital.com/insights/multi-agent-systems 2026-08-17T00:00:00.000Z The pitch for a swarm is that thirty agents bring thirty approaches. Anthropic ran the experiment. Eighteen of thirty agents created the same git branch, with the same name. Build an LLM wiki, not another chatbot https://salamoncapital.com/insights/llm-wiki 2026-08-17T00:00:00.000Z The cheapest upgrade available to most companies is not a better model. It is a folder of markdown files that nobody has written yet. Context engineering is subtraction https://salamoncapital.com/insights/context-engineering 2026-08-17T00:00:00.000Z The instinct is to give the model more information. The research says every token you add makes it worse at using the ones that mattered. The pipeline, not the headcount https://salamoncapital.com/insights/ai-workforce-transformation 2026-08-17T00:00:00.000Z The boardroom question is how many jobs go. Anthropic surveyed 9,700 people who actually use this at work, and the exposure sits somewhere almost nobody is guarding. The approve button is not a control https://salamoncapital.com/insights/ai-agent-guardrails 2026-08-17T00:00:00.000Z Every team building an agent reaches for the same safeguard: ask the human first. Anthropic measured what humans actually do with those prompts. They approve 93% of them. What an AI agent actually costs https://salamoncapital.com/insights/ai-agent-cost 2026-08-17T00:00:00.000Z Everyone compares the price per million tokens. It is the one number on the invoice that decides the least. Which work to hand to an AI agent first https://salamoncapital.com/insights/which-work-to-automate-first 2026-08-15T00:00:00.000Z The question that decides whether an AI project pays for itself is not which model to use. It is which piece of work goes first — and there is a defensible order.