# Gemini 3 Flash should NOT EXIST. TOP 2% Engineering. /PLAN 2026 Source: https://www.youtube.com/watch?v=u-SQ0Jsv4mI Recap page: https://rapidrecap.app/video/u-SQ0Jsv4mI Generated: 2025-12-22T14:39:18.422+00:00 --- ## Quick Overview The central thesis for achieving top 2% engineer status in 2026 revolves around increasing trust in agentic systems, which the speaker defines as the 'Year of Trust,' enabling greater speed, iteration, and impact through advanced agentic workflows like custom agents and multi-agent orchestration. **Key Points:** - Gemini 3 Flash is described as an 'absolutely insane' model that 'should not exist' because it offers top-three intelligence at a top-five price, with usage on Open Router showing a 385% token increase preview usage. - The main limitation for engineers is no longer the models themselves, but the engineer's ability to trust their agents, making 2026 the 'Year of Trust,' where trust equals speed, speed equals iteration, and iteration equals impact. - Top 2% engineers will continue to bet on Anthropic due to its consistent execution, best tooling for engineers (like Claude Code), and superior performance in long sequences of tool calls, which is the measure of trust. - Tool calling represents the greatest opportunity, as only 15% of output tokens are tool calls, which the speaker views as a rough proxy for impact; pushing agents to call longer chains of tools directly increases impact. - The most important bet is on building custom agents, as 'one well-crafted agent can replace thousands of hours of manual work' by solving custom problems for money, requiring expertise in prompting, context engineering, and SDK usage. - The next frontier after custom agents is multi-agent orchestration, where 'one agent is not enough,' requiring lead agents (command level agents) to delegate, monitor, and coordinate worker agents to increase compute and verification. - The speaker predicts the AGI hype will die as engineers focus on practical agent delivery, leading to the emergence of the first end-to-end agentic engineers who ship features 'prompt to production, no review, no human in the loop.' **Context:** Andy Dev Dan, who places himself in the top 10% of engineers using agentic technology, outlines his roadmap for becoming a top 2% agentic engineer by 2026, based on his accumulated work and conversations. This plan is structured around a central thesis for the upcoming year, which he identifies as the 'Year of Trust,' contrasting the current state of powerful foundational models like Gemini 3 Flash and Opus 4.5 with the necessary human element of trusting these automated systems to scale impact. ## Detailed Analysis The speaker asserts that the current model releases, including Gemini 3 Flash, Opus 4.5, and GBG 5.2, signal intense competition that drives down cost per intelligence, empowering engineers. However, the true bottleneck is the engineer's trust in their agents, framing 2026 as the 'Year of Trust' because trust directly translates to speed, iteration, and ultimately, impact. Key bets for top engineers include continuing to back Anthropic for its superior tooling and execution in agentic workflows, and maximizing tool calling, which is currently underutilized at only 15% of output tokens, as tool calls are a direct proxy for agent action and impact. The highest value lies in building custom agents to solve specific, profitable problems, as these agents, built with minimal code, can automate massive amounts of work, leading to skyrocketing trust when fine-tuned. Scaling this is achieved via multi-agent orchestration, moving beyond single-agent execution to command-level agents coordinating specialized worker agents, which increases verification and trust. To manage this scale safely, agents must run in isolated agent sandboxes, allowing engineers to defer trust until a result is ready for integration, mirroring traditional staging environments. Furthermore, top engineers will shift tasks out of slow in-loop terminal prompting into out-loop agent decoding systems, reserving manual intervention only for critical tasks. Public benchmarks will become saturated, forcing top engineers to rely on private evaluation systems specific to their use cases, while agents are predicted to consume Software as a Service (SaaS) applications by becoming the primary interface. Finally, the speaker rejects AGI hype, focusing instead on shipping tangible value with agents, culminating in the prediction of the first end-to-end agentic engineer emerging, who ships features from prompt to production without human review, representing the ultimate expression of trust in engineered systems. ### Model Landscape & Economics - Gemini 3 Flash is 'insane' due to top-three intelligence at top-five price - Intelligence per token is increasing, meaning cost per intelligence is decreasing - Competition ensures optionality for engineers. ### Central Thesis for 2026 - The year is defined by 'Trust' in agents, as trust equals speed, speed equals iteration, and iteration equals impact - The limitation is human inability to push agents further due to lack of trust. ### Anthropic & Tool Calling Opportunity - Top engineers will bet on Anthropic for best tooling, tool calling, and consistent execution - Tool calling, a proxy for impact, is severely underutilized at only 15% of output tokens. ### The Power of Custom Agents - Custom agents solve custom problems that generate money - One well-crafted agent (50 lines of code, 3 tools, 150-line prompt) can automate thousands of hours of manual work. ### Scaling with Orchestration and Safety - 'One agent is not enough'; Multi-agent orchestration, led by a command-level agent, is the next big trend - Agents must run in isolated agent sandboxes to defer trust and provide a safe staging environment for scale. ### Workflow Evolution - Top engineers will progressively offload tasks to out-loop agent decoding systems, minimizing slow, in-loop terminal babysitting - Agents are eating SaaS, as UIs are replaced by prompt interfaces driven by multi-agent backends. ### Predictions Track Record & Future Bets - Speaker correctly predicted AI coding becoming standard and agentic coding beginning, but incorrectly predicted OpenAI remaining number one and infinite memory being solved - Big Bet 11 predicts the first end-to-end agentic engineer shipping features prompt to production without human review.