Apuntes de ingeniería desde el terreno.
Ensayos y notas de campo sobre IA, ingeniería de software, diseño y el oficio de construir equipos de producto que entregan. Escrito por los ingenieros que hacen el trabajo. Publicaciones en inglés.
Anthropic Just Published the 2026 Agentic Coding Trends Report — Engineers Are Using AI in 60% of Their Work but Fully Delegating Only 0–20% of Tasks, Rakuten Ran a Single Seven-Hour Agent Across a 12.5 Million-Line Codebase, the Engineer's Role Has Moved From Implementer to Orchestrator, and the Highest-Adoption Cohort Is Staff+ Engineers — The Coding Agent Is a Senior Amplifier, Not a Junior Substitute, and Every Engineering Team's FY27 Plan Has to Land Against the New Topology.
Anthropic published the 2026 Agentic Coding Trends Report in June 2026, drawing on telemetry across the Claude Code enterprise install base, customer interviews, and a structured questionnaire that ran through April and May. The report identifies eight trends reshaping how software gets built; two are load-bearing for every engineering team's FY27 plan. First: engineers report using AI in roughly 60% of their work but report being able to fully delegate only 0–20% of tasks — the gap between AI-shows-up-in-most-of-my-day and AI-runs-the-task-end-to-end is the operating reality the discipline has to be designed around. Second: the engineer's role is moving from implementer to orchestrator, with value shifting toward system design, agent coordination, quality evaluation, and senior-judgment review. The headline practitioner case is Rakuten running a single agent on a complex feature implementation across a 12.5 million-line codebase in a seven-hour autonomous run that landed merged code. The multi-agent coordination pattern is now the dominant pattern in the highest-productivity install-base cohort. And the highest-adoption cohort is staff+ engineers — the coding agent is a senior amplifier, not a junior substitute. The structural read isn't 'AI coding tools are getting better.' It's that the shape of the engineering team that uses them is changing. Here's what the role shift restructures about how teams ship code, where the report is signal and where it is over-extrapolated, and the senior-judgment work that turns the report's framing into a compounding productivity delta.
GitHub Copilot Just Replaced Premium Requests With Per-Token AI Credits on June 1 — Developers Watched 10x to 100x Cost Swings Land on Their Dashboards on Day One, 82% of a Monthly Allowance Burned in a Single Editor Session, and Every Coding-Agent Buyer's Procurement Conversation Just Acquired a Per-Token Accountability Layer — The FinOps Discipline That Has Been a 'We'll Look at It Next Quarter' Item Is Now a Live Monitoring Surface, and the Engineering Team's Senior Judgment on Routing Just Got the Receipts to Match.
On June 1, 2026, GitHub flipped every Copilot plan from the prior premium-request allowance to usage-based billing through AI Credits — 1 credit = $0.01, with input, output, and cached tokens all metered against the published per-model API rates. Copilot Pro is now $10/month with $15 in credits; Pro+ is $39/$70; Business stays at $19/$19; Enterprise stays at $39/$39; the new Max tier lands at $100/$200. Monthly-plan holders migrated automatically; annual-plan holders stay on the legacy allowance until renewal. Day-one developer reports were sharp: 10x to 100x cost swings on heavy workloads, one engineer watching 82% of a monthly allowance evaporate in a single editor session, the long tail of teams discovering they had been quietly running the most expensive flagship behind every tab-complete keystroke. The structural read isn't 'the IDE vendor raised prices.' It's that the cost surface of coding-agent usage — a flat, predictable, $19/seat line item on the procurement spreadsheet — just became a per-token accountability layer the engineering team has to monitor like any other production cost. The model-selection decision the IDE vendor was quietly subsidizing is now the buyer's routing decision. Here's what the meter restructures about coding-agent procurement, where the per-token number is signal and where it is noise, what the engineering team owes the discipline in the first 90 days, and the senior-judgment work that turns the meter into a compounding cost-and-quality advantage instead of a reactive panic line item.
Google Just Killed an Open-Source CLI on a 30-Day Notice and Replaced It with a Closed-Source Binary — Gemini CLI Stops Serving Free, Pro, and Ultra Requests on June 18, the Migration Target Is the Closed-Source, Go-Based Antigravity CLI With a New Plugin Surface and Asynchronous Multi-Agent Orchestration, and Enterprise License Holders Are Carved Out — The Tooling-Lock-In Risk Just Got a Concrete Case Study, and Every Engineering Team's Coding-Agent Stack Has to Be Re-Audited.
On May 19, 2026, Google announced that Gemini CLI — the open-source command-line agent the company shipped at Google I/O 2025 — would stop serving requests on June 18, 2026 for Google AI Pro, Ultra, and free-tier users. The migration target is Antigravity CLI: a closed-source, Go-based binary (agy) that shares the agent harness with the Antigravity 2.0 desktop application, ships asynchronous multi-agent orchestration, and replaces the prior Agent Skills, Hooks, Subagents, and Extensions surface with a new plugin model. Enterprise Gemini Code Assist Standard and Enterprise license holders retain access to the prior CLI indefinitely; everyone else has thirty days to migrate, retest their automation, re-author their hooks, and re-implement any extension that does not survive the plugin-model transition. Google's stated rationale — 'your workflows have simply outgrown those early days of 2025' and 'users require multiple agents communicating with each other' — frames the move as a feature upgrade. The honest read is harder: an open-source CLI with a public source tree, a community PR pipeline, and the broader open-source default-tier hygiene that came with it is being deprecated in favor of a closed-source binary with a new license, a new plugin model, and an enterprise-only access carve-out that splits the user base. Here's what the forced migration teaches every team about tooling lock-in risk on coding agents, the audit the engineering team owes its current coding-agent stack against the OpenCode-style model-agnostic default, the migration discipline the team has thirty days to execute, and the senior-judgment work that determines whether the next coding-agent decision survives the next vendor's release cycle.
OpenCode Just Took the #1 Spot on the AI Coding-Agent Rankings — 160,000 GitHub Stars in Under a Year, 7.5 Million Monthly Developers, 75+ Provider Integrations, LSP Auto-Loading, Multi-Session Subagents, and Zero IDE Lock-In — The Model-Agnostic Open-Source Default Just Dethroned Cursor on LogRocket's June Power Rankings, and the Engineering Team's Coding-Agent Procurement Conversation Just Reset Around Vendor-Neutral by Default.
OpenCode — the model-agnostic, terminal-first, open-source coding agent from the SST team — crossed 160,000 GitHub stars in under twelve months from its June 2025 launch, climbed to 7.5 million monthly developers, and took the #1 spot on LogRocket's June 2026 AI dev tool power rankings, displacing Cursor in the most consequential single ranking shift the category has seen since Cursor 3's rebuild. The architecture choice that drove the growth is the same architecture choice that reshapes every team's coding-agent procurement decision for FY27. OpenCode treats the LLM as a pluggable dependency — 75+ providers including Claude, GPT, Gemini, the open-weight frontier coding tier, and local models routed through the same uniform agent surface — with native LSP auto-loading per language, shareable sessions, multi-session subagents running in parallel against the same monorepo, GitHub Copilot and ChatGPT Plus/Pro account authentication, and a privacy posture that does not store the team's code or context. The structural read isn't 'a popular open-source tool overtook a popular closed-source tool.' It's that the model-vendor lock-in conversation that has been the silent procurement axis on every coding-agent decision since 2023 just got a credible, well-engineered, large-installed-base default that puts the model decision back where it belongs — at the call site, graded against the workload — instead of at the IDE-vendor commercial relationship. Here's what the model-agnostic default does to the team's coding-agent stack, the routing-table discipline the model-agnostic surface unlocks, the eval-and-monitoring surface every engineering org now owns, and the senior-judgment work that turns the substrate into compounding production capability.
Anthropic Just Put a Daily-Updated Scoreboard Behind Every Claude Implementation Vendor — The Services Track and Partner Hub Land on Top of a $100M Fund With Three Public Tiers, 10,000 Certified Consultants, 40,000 Firms in the Pipeline, and an MCP Connector That Lets a Buyer Ask Claude Itself Who's Qualified — The Procurement Conversation for AI Implementation Just Got an Anthropic-Curated Default Surface, and the Vendor-Neutral Boutique Has to Decide Where It Stands.
On June 3, 2026, Anthropic shipped the Services Track and Partner Hub on top of the Claude Partner Network — the $100 million fund the company committed in March 2026 to certify, train, and co-market the implementation firms that ship Claude into enterprise production. The Services Track formalizes three publicly visible tiers — Select (10 certified practitioners, 2 production deployments, 1 customer story), Preferred (100 / 15 / 3), and Global Premier (1,000 certified practitioners, 100 deployments across 3+ regions, 15 stories, jointly developed business plan with named Anthropic executive sponsors). The Partner Hub is the publicly visible directory enterprise buyers use to find qualified partners, with a daily-updated dashboard showing where each firm stands against the published thresholds and an MCP connector that lets a buyer ask Claude itself about a partner's standing. Over 10,000 consultants already hold Claude certifications through the Anthropic Partner Academy and over 40,000 firms have applied to the program. The structural read isn't 'Anthropic added a directory.' It's that the procurement default for Claude implementation services just acquired an Anthropic-curated, daily-graded, publicly auditable scoreboard that the buyer can use to short-list vendors without doing their own research — and that the vendor-neutral boutique that isn't pursuing tier promotion has to articulate its position against a public scoreboard rather than against an invisible one. Here's what the program restructures about the implementation-services procurement decision, where the certification ladder is signal and where it's noise, what the buyer should weigh on top of the tier number, and the engineering and human-judgment discipline that determines whether the implementation team is calibrated against the workload or against the certification ladder.
Stanford's AI Index 2026 Landed — Junior-Developer Employment Down 20%, Entry-Level Software Postings Down 67% Since 2023, Senior Roles Up 6–12% in the Highest AI-Exposure Categories, and the Productivity Gain Concentrating Asymmetrically at the Senior Tier — The Engineering Team-Shape Decision Is Now the FY27 Budget's Most Consequential Procurement Call.
The Stanford 2026 AI Index Report crystallizes the labor-market shift that's been moving for eighteen months. Employment among software developers aged 22–25 fell nearly 20% from its late-2022 peak by July 2025, while workers 30 and older in the highest AI-exposure occupations grew 6%–12% over the same window. Entry-level software-engineering postings in the U.S. dropped 67% between 2023 and 2024 according to Stanford's Digital Economy Lab analysis. Indeed Hiring Lab reports senior tech job titles down 19% vs five years earlier and junior titles down 34%. Big-tech entry-level hiring dropped more than 50% over the last three years. The 26% measured productivity gain from AI tools in software engineering is the population average — concentrated at the senior tier, meaningfully weaker at the junior tier, and decisively asymmetric in its team-shape consequences. The structural read is uncomfortable but coherent: a team of five senior engineers with AI tools is now doing what previously required a team of eight, and the three positions removed from the math are disproportionately the junior-developer slots. Here's what that does to the senior-engineer hiring market, the internal training pipeline that used to convert juniors to seniors over four-to-six years, the in-house senior:engineer ratio every team has to reconsider for FY27, and the procurement decision around senior-only external capacity that's a meaningfully different conversation in 2026 than it was in 2023.