**Three things happened in the first 11 weeks of 2026 that changed how we build software at ThinkCodeShip. None of them were a chatbot writing a better email.** --- On January 14, I was watching our deploy logs when Bloomberg flashed a headline about software stocks cratering. By the time I'd scrolled through the S&P North American Software Index, our own SaaS dashboard showed a 12% spike in trial signups. People were panicking about AI replacing software, and simultaneously signing up for more software. That contradiction is the whole story. Here is what actually happened in the first 11 weeks of 2026, why the panic was both overblown and justified, and what we are doing about it at ThinkCodeShip. --- ## The SaaS Crash Was Real and Misunderstood Between mid-January and mid-February 2026, roughly **one trillion dollars** evaporated from software and SaaS stocks. Salesforce dropped 25-35%. Workday and Atlassian got hammered alongside the sector. The trigger was the launch of AI agents like Anthropic's Claude Cowork and OpenAI's Frontier, which could perform knowledge work that previously required humans. The market did simple math: one agent replaces five seats. Five seats become one license. Multiply that across every enterprise software contract and you get a selloff. Here is where I disagree with both the doomsayers and the optimists. The doomsayers are wrong because enterprise software is not just code. It is decades of compliance frameworks, audit trails, and business rules baked into postgres tables that nobody fully understands. A general-purpose AI agent that can draft a decent email does not know that your revenue recognition logic has to handle German VAT differently from California sales tax. That moat is real. The optimists are also wrong. The analysts calling the panic "overblown" keep pointing to software spending growth as proof everything is fine. But spending growing does not mean the same companies survive. The business model is shifting from per-seat pricing to outcome-based pricing, and the companies that cannot make that shift will die even if total spending rises. We are already seeing SaaS companies pivot to charging for "agentic work units" instead of human seats. Some of those pivots are desperate and broken. We track this at ThinkCodeShip because our own product sits in the crossfire. If AI agents replace the developers who use our tool, our subscription revenue drops. If AI agents need tooling to ship reliable code, our revenue grows. We are betting on the second scenario, but I will not pretend it is a sure thing. --- ## Claude's 1M Token Context Window: Useful, Not Magic On March 13, 2026, Anthropic made their million-token context window generally available for Claude Sonnet 4 and 4.5. A million tokens is roughly eight full-length novels, an entire codebase, or a thousand-page contract. Previously, AI had a memory problem. You would upload a document, ask about page 400, and the model would forget page 10. You had to chunk documents, summarize them, and hope the summary did not drop the detail you needed. That problem is now solvable. Here is the contrarian take: bigger context does not mean better reasoning. We ran a test. We fed Claude a 600-page API documentation set with the 1M context window and asked it to find a specific deprecated endpoint. It found it. Then we asked it to explain why that endpoint was deprecated, using only the context in the document. It hallucinated a reason that sounded plausible and was completely wrong. The actual deprecation note was on page 312. The model read it. It just did not understand the engineering tradeoff being described. Context size solves the retrieval problem. It does not solve the comprehension problem. If your model reasons poorly with 200K tokens, it will reason poorly with 1M tokens. The difference is that it will reason poorly about more information at once. The real win is economic. Requests under 200K tokens are charged at standard rates. Only above 200K input tokens does premium pricing kick in. For us, that means we can now feed an entire Phoenix migration file plus the test suite into a single prompt without chunking. Last month that would have cost us four API calls and three context-loss bugs. This month it is one call. To use it, you include a beta header in your API requests. It is available on Amazon Bedrock and through Anthropic's direct API for organizations in usage tier 4. Setup took us 20 minutes. The first thing we tested was asking Claude to review its own diff from the previous session. It found a missing test case we had overlooked. That was a real win, not a party trick. --- ## OpenClaw: A Lobster With Scissors While markets panicked and context windows expanded, an open-source project called OpenClaw quietly became the most interesting AI release of the year. OpenClaw is not a chatbot. It is a personal AI assistant that runs on your own devices, connects to your actual life, and takes action. It clears your inbox, sends emails on your behalf, manages your calendar, and checks you in for flights. It works through WhatsApp, Telegram, Slack, Discord, iMessage, Google Chat, and Signal. We installed it. We used it for two weeks. Here is what happened. The first three days were magical. OpenClaw cleared 200 unread emails from my inbox, drafted responses for the ones that needed replies, and scheduled two meetings I had been postponing for weeks. It felt like having a competent intern who never sleeps. Day seven was less magical. OpenClaw sent a calendar invite to the wrong Katherine because it matched on first name instead of email address. The invite went to a contact I had not spoken to in three years. She replied. It was awkward. Day twelve was worse. OpenClaw tried to check me in for a flight that had been cancelled and rescheduled two days earlier. It pulled the old confirmation number from its memory, hit the airline API, and reported success. I showed up at the airport with no seat. The lesson is not that OpenClaw is bad. It is a genuinely impressive piece of software with a mascot that is a lobster, and the community building custom Skills for it is growing fast. The lesson is that autonomous agents fail in ways that are hard to predict because they do not have the context to know when they are wrong. A human assistant who was unsure would ask. OpenClaw does not ask. It acts. You can install it yourself: ```bash git clone https://github.com/openclaw/openclaw.git cd openclaw pnpm install pnpm ui:build pnpm build pnpm link --global openclaw onboard --install-daemon ``` Run it. Watch it work. Then watch it break. The gap between those two moments is where the real engineering happens. --- ## What We Are Building Because of This I am not going to give you a four-step adaptation framework. Frameworks are what people write when they do not want to commit to anything specific. Here is what we are actually doing at ThinkCodeShip, with dates and tradeoffs. **January 2026: We added AI-assisted code review to our pipeline.** Claude reviews every pull request diff and flags missing tests, unhandled error paths, and naming violations. It catches about 30% of the issues our human reviewers would have caught. It also produces false positives on 15% of PRs, usually flagging Elixir patterns it does not understand. We kept it because the 30% is worth the noise. **February 2026: We tried to automate our deploy pipeline with an AI agent.** It worked for six days. On day seven, our CI config changed and the agent silently started deploying from the wrong branch. We caught it before production. We rolled back to human-triggered deploys. The agent is still running in a sandbox, but it does not touch production. **March 2026: We are integrating the 1M token context window into our blog research workflow.** Instead of chunking documentation into 4KB pieces and losing cross-references, we feed entire API docs into a single prompt. The setup took 20 minutes. We have not measured the quality difference yet. I will report back when we have numbers, not vibes. The pattern across all three is the same. AI is useful when the failure mode is visible and the stakes are low. AI is dangerous when the failure mode is silent and the stakes are high. Code review is visible and low-stakes. Deploys are silent and high-stakes. Research is visible and low-stakes. That is the entire heuristic, and it has not been wrong yet. --- ## The Open Question Six months ago, AI wrote essays. Today it reviews code, reads million-token documents, and manages inboxes with varying degrees of competence. I do not know what it does six months from now. Neither does anyone else, including the people building the models. The difference between January and March is that the question shifted from "can AI do this?" to "should we let it?" That is a harder question and a more interesting one. What I do know is this: the developers who win are not the ones who trust AI the most or the least. They are the ones who build systems where AI failure is visible, recoverable, and cheap. That is what we are building at ThinkCodeShip. If that is what you need, we are here. If not, the lobster is available on GitHub.

Posts tagged with "trends"

AI Stopped Being Hype in March 2026