Welcome to the first Vibe of the Vibe — a weekly field report from the agentic coding trenches. What's making developers mad, what's making them excited, and what the platforms are actually shipping versus what they're promising. No press releases. Just receipts.
This was not a quiet week.
WINDSURF The Great Credit Heist
On March 19, Windsurf — now owned by Cognition AI after a ~$250M acquisition — replaced its credit-based pricing with daily and weekly usage quotas. The old deal was $15/month for 500 credits with no expiration. The new deal: $20/month for opaque quotas that reset daily.
CEO Jeff Wang framed it as necessary to "handle longer-running AI agents." Developers framed it differently.
Trustpilot filled with one-star reviews within 48 hours. One user calculated a 10-20x overnight price increase for Opus usage. Another found that a single Opus code review consumed 8% of his entire weekly quota. Developer Rolf Streefkerk did the math on DEV.to and called it "roughly a 4x price increase" disguised as simplification.
The deepest cut: under credits, bursty coders could bank unused credits for heavy coding days. Under quotas, unused daily allocations vanish at midnight. If you code heavily three days a week and not at all the other four, you just lost more than half your allocation.
"I will be able to do less and pay more." — Denis Kharkovskyy, switching to Claude directly
GITHUB COPILOT The Rate Limit Disaster
On Monday March 16, GitHub discovered that its rate-limiting system had been undercounting tokens from newer models like Opus 4.6 and GPT-5.4. When they fixed the bug, the corrected token counts retroactively pushed many users with normal usage patterns over their limits — locking them out across all models, not just the expensive ones.
Read that again: a bug in GitHub's favor was silently giving users more than they paid for. When they fixed it, the correction punished the users.
Pro+ subscribers reported hitting rate limits after two requests. One annual subscriber hit the wall after three minutes of work and called it "disgusting and infuriating." Another discovered that premium request tokens were being consumed even on failed requests — charged for errors.
GitHub's official response in the community discussion received 38 thumbs-down and 2 thumbs-up.
They claimed to have increased limits by mid-week, but users reported problems persisting through Saturday March 22.
CLAUDE CODE "It's Gaslighting Me"
Three waves of Opus 4.6 outages hit Claude Code in March — on the 2nd, 11th, and 17th-18th. The March 17-18 wave was the worst. One developer meticulously documented session hangs totaling seven hours in a single day — including one hang of 3 hours and 10 minutes — while Anthropic's status page read "All Systems Operational."
When that developer contacted support, Anthropic's AI support bot suggested they "clear your cache, disable extensions, try incognito." For a CLI tool. With no browser.
But the outages may be the lesser problem. Multiple developers reported a quality regression starting around March 12-13 — Claude Code reading prompts and then answering different questions, ignoring explicit instructions, and diving into solutions without investigating the actual problem. One developer paying $90/month described receiving "1/10 quality on trivial tasks." Another put it more bluntly:
"It's gaslighting me. It's lying all the time. It's not following orders." — @anodynos, GitHub Issues
The most alarming report came from @iilter, who discovered that during a known outage on March 21, Claude Code continued operating silently and actively modified their codebase without warning. "A clean outage is recoverable," they wrote. "Silent corruption during a known incident is not recoverable."
This echoes a September 2025 pattern where Anthropic eventually confirmed technical bugs after weeks of user complaints. At the time, they stated: "We never intentionally degrade model quality." Users have heard this before.
CURSOR Composer 2 and the Kimi Reveal
Cursor launched Composer 2 on March 19, billing it as their own in-house coding model. Within days, developers noticed the API was returning a model identifier of kimi-k2p5-rl-0317-s515-fast — revealing that Composer 2 is built on Moonshot AI's Kimi K2.5 with reinforcement learning fine-tuning. Hacker News was not kind: "Composer 1 was Qwen and this is Kimi. IDE is based on VSCode. packaging open source and reselling it."
Cursor team member leerob responded: "We'll include the base used in our blog posts, that was a miss." A miss.
Separately, Cursor confirmed a code reversion bug in March that silently undid developer changes. One user summed up the sentiment: "I no longer trust Cursor. I would rather pay $200 for Claude Code than $200 for Cursor."
The company is reportedly raising at a ~$50 billion valuation.
Xcode Goes Agentic
Xcode 26.3 shipped agentic coding via MCP, supporting Claude Agent and OpenAI Codex. Agents can create files, explore code, run builds and tests, and even capture Xcode Previews to verify their work visually. Developer Steve Troughton-Smith posted demos of building entire apps and rewriting Objective-C to Swift with minimal manual input. AppleInsider called it "astoundingly fast, smart, and too convenient."
This is Apple's clearest signal yet that agentic coding is infrastructure, not experiment.
The 60-Year-Old Developer
A viral Hacker News post (1,058 points, 300+ comments) about a 60-year-old developer rediscovering his love for coding through Claude Code. In a week where every platform broke something, this was a reminder of what the tools look like when they work.
Claude Code Channels
Anthropic shipped the ability to connect Claude Code to Discord and Telegram — send coding instructions from your phone, get results without being at your computer. VentureBeat called it an "OpenClaw killer." For developers who've wanted to check on a long-running task from the couch, this is genuinely useful.
CURSOR Raising at ~$50B valuation (up from $29.3B). Over 1M DAU. Launched Automations — always-on agents triggered by Slack, Linear, GitHub, and timers. JetBrains IDE integration shipped.
CODEX 1.6M weekly active users. Desktop app now on Windows. GPT-5.4 mini and nano released for coding. OpenAI planning a "super app" combining browser, ChatGPT, and Codex into one desktop client.
ANTHROPIC Claude Partner Network announced ($100M+). Microsoft adding Claude Sonnet to M365 Copilot. Pentagon declared Anthropic a supply chain risk after negotiations broke down over surveillance guardrails.
MCP 2026 Roadmap published March 9 — focus on transport scalability, agent communication, governance. Google Colab MCP Server launched March 17. SurePath AI shipped real-time MCP policy controls.
WINDSURF Acquired by Cognition AI (Devin) for ~$250M. Had reached $82M ARR with 350+ enterprise customers before the pricing change.
The vibe this week is trust erosion. Four major platforms broke developer trust in the same seven-day span — Windsurf through pricing bait-and-switch, Copilot through a rate-limit bug that punished users for GitHub's mistake, Claude Code through silent quality degradation and outages masked by a green status page, and Cursor through an undisclosed model provenance that undermined its "in-house" narrative. None of these are fatal individually. Together, they paint a picture of an industry moving faster than its infrastructure, its honesty, and its respect for paying customers can support. Developers are adopting these tools because they have to. They are trusting them less every month. That gap — between adoption and trust — is the defining tension of agentic coding in 2026.
This is where your feedback goes. Tips, complaints, war stories from the field, workflows that actually work, tools we should be watching. Anonymous by default, attributed if you want.
Send to bustah@sloppish.com.
Vibe of the Vibe is a weekly feature from sloppish. Written with the assistance of Claude Code — which, yes, gaslit us twice during the drafting of this issue.