From my perspective, I see 8 pillars emerging, each supporting the other in a sort of octagon or octet truss. I’ve spent the most time thinking about/experimenting with 1-3:

1. Accelerated Creation

In software development Agentic IDEs are the latest wave following vibe coding. I think multi-agent workflows will continue to develop and we’ll begin to see competing multi-agent design patterns (e.g. planner+coder+reviewer+tester vs rad-coder+refactorer+tester+tester/reviewer) and templates for different patterns.

My thoughts on tools after ~6 months of usage:

• Cursor – leads the pack on features and best automatic utilization especially if you’re using different agents with different frontier models, unfortunately they have made it difficult to use with a local model on your machine without some hacking and proxying which I think could be their long-term downfall but we’ll see how they and other competitions adapt. Pricing using Anthropic models for me is also ironically better using Cursor than Claude Code, not sure why but seems to be some optimizations and indexing cursor is doing under the hood with conversation context.

• Claude Code CLI – I would describe this experience as most like traditional development in that you see and approve the code as it is being generated which is great for quality control. Honestly I’m torn between whether I like cursor or claude code more for individual development, but cursor edges out when it comes to multi-agent especially if I’m working on multiple APIs at once in parallel because I can easily kick off 4-5 agents in the IDE with those tasks then come back to review. The question is whether that is ultimately more performant than if I did it serially since I have to review each anyway, but theoretically if I had multiple reviewers I trusted than that bottleneck would go away for cursor and clearly edge it out over claude.

• Aider – I use this like a specialized tool for the “code diff” workflow approach when I have a bug or surgical change. Generates 30-50% lower context usage which reduces my cursor/claude bill by almost 50%, allows me to assess larger combinations of context and is faster to iterate vs full hands off agentic

• VS Code with “Continue” – a bit of a “cursor light” but if you have a $500-$1500 graphics card and comfortable with a little more iteration using Qwen/Devstral

• Kiro – very opinionated goal oriented workflows which seem to be designed from a product first, hands off strategy. Can streamline development of common patterns once identified, such as iterating on and existing CRUDs services/APIs but this approach does not scale with broader context.

Still experimenting with Antigravity, Cline and Openclaw. My goal is to narrow it down to top 2-3 on instinct, then work on a benchmark for assessing productivity gains.

2. Autonomous Agents

In my mind this is like “the new REST”. Agent to Agent Integration/Communication (currently primarily through MCP) will continue to become multilayered and multifaceted. My agent will call your agent who will call the restaurant’s agent and collectively they will find us a nice spot for dinner sometime and book it. I think google is currently the leader here with A2A but there is alot of open source development also:

• https://developers.googleblog.com/en/a2a-a-new-era-of-agent-interoperability/

• For gdocs, calendar etc – https://github.com/google/mcp

• Jira and Confluence – https://github.com/atlassian/atlassian-mcp-server

3. Agentic Orchestration and MLOps

If autonomous agents are “the new REST” then agentic orchestration/MLOps is the new distributed systems. This is less about how the agents communicate and more about the architecture of autonomous agents and how they retrieve their data, authenticate, use different models to be able to fulfill their agent/service responsibilities etc.

This article has nice coverage that I agree with:
https://medium.com/@anil.jain.baba/top-5-agentic-ai-frameworks-tools-and-services-fb51d5876154

4. Model proliferation

This might be at the core of the octet truss model I’m conceptualizing and at this point in 2026 model proliferation probably speaks for itself. Frontier/foundation models lead the way with broad open source experimentation creating a feedback loop to big tech and things like Yann Lecun’s “world models” coming soon. This is primarily a research space but we’ve only scratched the surface on industry specific expert systems for things like autonomous machinery/self-driving and medical/chemical applications which will open up a new wave of B2B/B2C SaaS apps.

My thoughts on current frontier models: 

• Anthropic/Opus/Sonnet – best balance on effort/prompting clarity to adherence and results with speed, I think their models dominate when working on pure engineering problems but they are not as good as processing product documents/jira tickets and business requirements etc

• ChatGPT – a bit converse to anthropics, much better at parsing the business stuff but worse at the code generation, the code tends to be less succinct and sometimes overcomplicated which is a general problem with all the models but anthropic’s do the best

• Gemini – somewhere between anthropics and chatGPTs models IMO. better at coding/technical than chatgpt but not as good as opus/sonnet, better than opus/sonnet at biz reqs/product docs/tickets but not as good as chatgpt

• Qwen – The remote models are comparable to gemini IMO and basically a less expensive gemini clone. The nice thing about these for me is I can run 2.5-coder and 3-coder-fast locally  on my GPU along with devstral, using qwen for coding and planning and devstral/mistral replacing chatgpt locally.

• Deepseek – I see it as a chatGPT variant that is a bit looser/less strict. If I’m probing for security and things that other models are overly sensitive about, deepseek usually doesn’t mind answering.

• Devstral/Mistral – for being open/”free” models they are very good and I can get similar results locally as the top frontier models but takes a little more meticulous and detailed prompting. When running locally, I use dual agents running each one and together they produce about the same result as chatGPT for analyzing biz/product docs to architecture but I still run chatGPT on the same problems in parallel while I’m experimenting and sometimes it edges out but surprisingly close to these open source models and chatGPT deosn’t edge it out as often as I would think for being a “frontier model”.

• Grok – in my local AI experimenting app I did also plugged in grok APIs but oddly enough their conversation and image generation over API seems to be way worse than when you are using the actual grok app or tools. I see 30% less prompt adherence than chatGPT/ takes more iterations to get the same results, but it is 3-4x cheaper/faster so there is that to be said about cost/speed at scale also so we’ll have to see how the industry adapts.

Many of these models have “thinking” versions and MoE versions which muddies the comparisons a bit but I still hold the above broadly true.

Notable mention to HuggingFace if you haven’t already been using it and you have a GPU, this place is the github of AI models.

5. Data Engineering/Optimization of data/metadata/tagging

I haven’t done much research here, but companies like Snowflake and Databricks appear to be at the forefront. One of the growing challenges which models are facing is a search problem – as context windows grow how do you determine what is relevant to store in the context relative to the prompt/query? I believe this subsearch problem will introduce new data structures and indexing patterns to optimize this challenge and represents a fundamental kind of O(n)->log(n) type of improvement in the way models consume, assess and iterate on data.

6-8. Inference optimization, Human-AI interface, AI governance/reliability/control

Not to diminish any one of these, but I am not diving into them. Inference optimization is more of a hardware problem and even software optimizations (e.g. in threading, CUDA optimization etc) will require cross-functional low-level systems expertise. Human-AI interface is the future of Ops and non-engineering teams along with AI governance/reliability/control which will influence how we develop but are more on the edge of software engineering IMO.

The Future

A “bubble” is possible in the same way the housing market and any other markets have bubbles. The notion is somewhat shortsighted to the impending leaps across all industries. Markets, like empires, rise and fall. The rate of growth of technology may fluctuate with global economics and affects the timeline of innovation, but not the overall potential. The advances on the horizon will usher in a new age of humanity. Circumstances might get a little “MAD”. We are either drivers of this future or unwilling participants. Either way we must choose to adapt – and be the change we want to see!