Predictions are cheap, so here are eight specific enough to be wrong — and grounded in where 2025 actually left us.
1. Verification becomes the headline skill
Generation is solved; correctness is not. In 2026 the engineers in demand are the ones who can rigorously verify agent output — through tests, evals, and judgment. "I can review what an agent built" outpaces "I can prompt an agent."
2. The harness eats the prompt
"Prompt engineering" recedes as a standalone skill. What matters is the whole harness — tools, context, memory, guardrails. The job is designing the environment an agent works in, not crafting the perfect sentence.
3. Model routing becomes standard practice
Sending every request to one frontier model will look as quaint as running every query against an unindexed table. Routing each task to the right-sized model becomes a default cost discipline, not an optimization.
The teams that win in 2026 are not the ones with the biggest model. They are the ones who use the right model for each job.
4. MCP servers proliferate inside companies
With MCP now vendor-neutral infrastructure, internal teams start exposing their systems as MCP servers the way they once exposed REST APIs. "Is there an MCP server for that?" becomes a normal question.
5. Security catches up to agents — painfully
Prompt injection, tool abuse, and over-permissioned agents cause real incidents in 2026. The teams that treated the agent surface as a serious threat model early look prescient; the rest learn the hard way.
6. Spec-driven workflows go mainstream
As agents implement reliably from precise specs, the spec becomes the high-value artifact. More teams restructure around writing and reviewing specs while agents handle implementation.
7. Small and open models do more of the work
Cheap, capable, often open-weight models absorb the high-volume routine tasks. Frontier models stay reserved for genuine reasoning. The economics push more workloads down-tier than people expect.
8. The senior role tilts further toward architecture and taste
Less time typing implementations, more time deciding what to build, setting constraints, and judging quality. Seniority becomes about judgment and system design more than raw output — and junior paths have to be rethought around it.
The meta-prediction
The unifying theme behind all eight: the bottleneck has moved off raw capability and onto everything around it — verification, context, security, economics, and judgment. 2026 is less about smarter models and more about the discipline of building dependable systems around them.
Check back at year's end and hold me to these.