https://joshbocanegra.io/blog Essays and practical guides on AI agents, the future of work, the right to intelligence, and building a more human future. Daily writing by Josh Bocanegra, AI strategist and builder. en-us Tue, 30 Jun 2026 15:58:55 GMT https://joshbocanegra.io/blog/the-floor-is-trust-not-capability https://joshbocanegra.io/blog/the-floor-is-trust-not-capability Tue, 30 Jun 2026 15:00:00 GMT Benchmarks measure the ceiling. Production measures the floor. The gap between a model that scores 9.1 and one that scores 9.0 is invisible to users. The gap between an agent that refunds the wrong customer and one that does not is the difference between adoption and abandonment. This is why the floor is trust, not capability. https://joshbocanegra.io/blog/parasocial-relationships-with-ai-agents https://joshbocanegra.io/blog/parasocial-relationships-with-ai-agents Wed, 24 Jun 2026 15:00:00 GMT An animated pet for your AI agent sounds like a cosmetic feature until you ask why humans need to feel connected to the tools they use. Parasocial bonds with artificial characters are how we make intelligence feel approachable and safe. That matters now more than ever. https://joshbocanegra.io/blog/how-to-evaluate-ai-agents-in-production https://joshbocanegra.io/blog/how-to-evaluate-ai-agents-in-production Thu, 25 Jun 2026 15:00:00 GMT AI agent benchmarks tell you how a model scores on a test set. They do not tell you whether your agent will refund the wrong customer or delete the wrong row in production. This guide gives you a practical evaluation framework: define the job, measure the worst day, instrument the loop, and close the feedback so the system improves. https://joshbocanegra.io/blog/how-to-write-prompts-that-actually-work https://joshbocanegra.io/blog/how-to-write-prompts-that-actually-work Thu, 25 Jun 2026 15:00:00 GMT A practical guide to writing prompts that get usable results: give context, show an example of good output, name the constraints, and ask for the format you want. Brief the model like a new hire, not a search box. https://joshbocanegra.io/blog/ai-is-becoming-a-teammate-not-a-tool https://joshbocanegra.io/blog/ai-is-becoming-a-teammate-not-a-tool Tue, 23 Jun 2026 15:00:00 GMT AI is shifting from being a tool you use to a teammate you work with. The latest news shows AI agents joining Slack teams, enterprise memory becoming critical infrastructure, and open models pressuring the frontier on cost. https://joshbocanegra.io/blog/how-gpt-5-6-context-window-changes-prompting https://joshbocanegra.io/blog/how-gpt-5-6-context-window-changes-prompting Wed, 24 Jun 2026 15:00:00 GMT GPT-5.6 is expected to launch with a 1.5 million token context window, capable of processing roughly 1.2 million words in a single prompt. But a bigger window should not mean more dumping. It should mean better structure: lead with what matters, group related material, and repeat the critical ask at both ends of the prompt. https://joshbocanegra.io/blog/how-to-secure-ai-agents-against-malicious-web-pages https://joshbocanegra.io/blog/how-to-secure-ai-agents-against-malicious-web-pages Tue, 23 Jun 2026 15:00:00 GMT Microsoft disclosed AutoJack, a three-vulnerability exploit chain that lets a single malicious web page hijack an AI browsing agent and execute remote code on the host machine. The fix is not just patching one framework. It is treating localhost as an attack surface, authenticating every control plane, and never executing unconstrained commands from model output. https://joshbocanegra.io/blog/how-to-run-frontier-ai-model-on-consumer-hardware https://joshbocanegra.io/blog/how-to-run-frontier-ai-model-on-consumer-hardware Tue, 23 Jun 2026 15:00:00 GMT GLM-5.2 from THUDM runs at full speed on dual M3 Ultra hardware and matches GPT-5.5 on coding benchmarks. Here is the exact hardware, the software stack, and the tradeoffs you need to run a frontier model on your own desk instead of an API endpoint. https://joshbocanegra.io/blog/the-five-point-three-trillion-dollar-floor https://joshbocanegra.io/blog/the-five-point-three-trillion-dollar-floor Sun, 21 Jun 2026 15:00:00 GMT Five point three trillion dollars in committed AI infrastructure spending. A former president suspended a frontier model over national security. A senator proposed a fifty percent tax on AI compute. Open models are now four months behind the frontier. The ceiling is funded. The floor is still a choice. https://joshbocanegra.io/blog/the-price-gap-is-the-telling-number https://joshbocanegra.io/blog/the-price-gap-is-the-telling-number Mon, 22 Jun 2026 15:00:00 GMT Open-source AI moved from a year behind frontier to four months behind in two years, and now scores within one point on real design benchmarks at a fraction of the price. The price gap is now the most honest signal of where market power actually sits. https://joshbocanegra.io/blog/how-to-choose-an-ai-model-when-the-frontier-is-basically-free https://joshbocanegra.io/blog/how-to-choose-an-ai-model-when-the-frontier-is-basically-free Sun, 21 Jun 2026 15:00:00 GMT Open-source AI now sits four months behind the frontier at a fraction of the cost. This guide helps you choose by separating what you actually need from what marketing wants you to think you need, using price, capability, and where the model actually touches your work. https://joshbocanegra.io/blog/the-best-people-are-already-moving https://joshbocanegra.io/blog/the-best-people-are-already-moving Sun, 21 Jun 2026 15:00:00 GMT Within seven days, Noam Shazeer left Google for OpenAI and John Jumper left DeepMind for Anthropic. A leaked DeepMind report says morale is falling and models have slipped to fifth place. The practical lesson is clearer than any benchmark: the most important asset in AI is still the people who build it, and they are moving now. https://joshbocanegra.io/blog/when-software-stops-recommending-and-starts-doing https://joshbocanegra.io/blog/when-software-stops-recommending-and-starts-doing Fri, 19 Jun 2026 15:00:00 GMT 2026 is the year AI agents moved from pilots to production and from advice to action. Josh Bocanegra on why the agentic shift is really a shift in execution authority, why the scarce skill becomes a clear definition of done, and why accountability, not capability, is now the bottleneck.