Where software value moves when agents eat the rest.
Most of what I’m watching in tech right now traces back to one thing: agents are eating the parts of software that used to need humans, dedicated tools, or both. Here are the shifts I’m tracking, and what each one changes.
Agents write most of the code now
Software engineers are leaning hard on agents. Tools like Claude Code and Codex now generate far more code than humans write by hand. That strains code review, maintainability, and testing. It also changes the shape of an engineering team: fewer people building more.
Agent harnesses replace dedicated tools
Tasks that used to need a dedicated tool now happen inside an agent harness. Demo videos with Claude Code and Remotion. User research by scraping Reddit. This raises a real question: should we build vertical agents or just skills and MCP servers that plug into a general agent?
Many agents, no shared orchestrator
AI-native teams run multiple agents in parallel across harnesses and platforms. It’s hard to keep track of them. We often don’t have enough visibility into what each one is doing. We need a Kanban board for agent work that spans platforms, with a built-in orchestrator.
Even small models find zero-days
Models are finding zero-day vulnerabilities in major operating systems and browsers, including bugs that sat in open source for years. Even small open-weight models can rediscover many of them. See Aisle’s writeup on AI cybersecurity after Mythos. Automated security checks on every change stop being optional.
RL saturates rewards
RL can saturate pretty much anything you can measure. Coding came first because tests give you a reward signal. Anything verifiable is next: engagement, investment returns, ad CTR. I don’t have a good feeling about RL-optimizing for dopamine, but it’s going to happen.
AutoResearch generalizes beyond ML
AutoResearch is a powerful idea: agents run experiments against a measurable objective. People use it for optimizing test-suite runtime, marketing experiments, and more. Anywhere you have a metric and code that can change that metric, you can point AutoResearch at it.
Search becomes agentic
Web traffic is shifting from search to AI assistants, and brands are scrambling for visibility inside them. OpenAI is already serving ads in ChatGPT. GEO/AEO is the new SEO.
Voice agents sound human now
Low-latency, interruptible voice models are good enough now. ElevenLabs voices are very realistic, even open local models can produce non-robotic speech. Agents will handle a lot of front-desk and customer-support work soon, with humans taking the trickier cases.
Software becomes a commodity
Agents make software faster and cheaper to build. The price floor drops. CRUD apps with little data transformation lose their high per-seat justification. But software-as-business doesn’t die. Products built on deep domain knowledge or layered on top of the commodity layer still command real prices.
MCP makes integration cheap
Integrating with external services used to be painful. With an MCP gateway and a catalog of MCP servers, you can wire thousands of services together with very little glue.
Context is the new bottleneck
Agents underperform because they’re missing the context humans have. That context lives in calls, Slack, code, email, and wikis. MCP gives an agent the connections, but the data still has to be queryable for the agent to use it. Glean is the clearest example: it indexes across your tools and gives agents a single place to query for the context they need.
The physical world is the new moat
If software is becoming a commodity, one durable moat is doing something in the physical world. That’s why a company like Figure is getting so much attention. Robotics is the one place an agent can’t beat you to the punch overnight.
Proprietary data is the other moat
Proprietary data is harder to bootstrap than prompting an agent. Wispr Flow improves transcription by learning from human edits. A GEO/AEO tool could aggregate ranking data across tenants to learn what actually works.
Some intelligence no longer needs the cloud
A growing set of AI tasks no longer needs the cloud. Apple already runs local inference on-device for summarization, rewriting, and speech recognition, and Ollama makes it trivial to run open models on your own machine. They’re nowhere near SOTA cloud APIs, but they’re good enough for plenty of work. The privacy and cost story is hard to beat.
Where the value moves next
If software gets cheaper to build and easier to operate via agents, value has to come from somewhere else. It’s moving to three places: the physical world (robotics), proprietary data (flywheels), and deep domain knowledge that the commodity layer can’t replicate.