This week a few things landed at once. Google made Gemini 3.6 Flash the default for its Managed Agents, betting on latency and reliability over raw capability. A new crop of orchestration tools shipped promising to move work from a to-do item straight to a finished pull request with no tool switching. Gartner put a number on the direction of travel: 40% of enterprise apps will have embedded agents by the end of the year, up from under 5% in 2025.
Read those together and a pattern shows up. The models are no longer the interesting part of the story. The plumbing around them is.
The bottleneck moved and most teams didn't notice
For two years the question was whether an agent could write the code. That question is mostly settled. Modern coding agents handle a scoped task, run for a while, and come back with something that compiles and often works. EY reported a 70% lift in software development productivity after rolling agent tooling out across tens of thousands of consultants. Those numbers are real.
But watch where an engineer's day actually goes now. It goes into two places the agent doesn't touch. First, getting a vague ticket into a shape an agent can act on: the acceptance criteria, the relevant files, the constraint that this service can't add a new dependency. Second, taking whatever the agent produced and deciding whether to trust it, which means reading the diff, checking the edge cases, and reconciling it with three other agents that touched nearby code the same afternoon.
That's the handoff. It sits on both sides of the agent, and it's where the time goes.
Why buying a better model doesn't fix it
Teams feel the friction and reach for the wrong lever. They swap Cursor for Claude Code, or upgrade to the newest flagship, and the ticket-to-PR cycle barely moves. The reason is simple. A smarter model writes better code from a good spec. It does nothing about a bad spec, and nothing about a review process that assumes one human reading one PR at a time.
Google's Flash-by-default move is a quiet admission of this. When you run an agent thousands of times a day, the marginal capability of the biggest model matters less than whether it responds fast and behaves predictably. Reliability is a workflow property, not a model benchmark.
Wire the handoff before you shop for tools
The teams pulling real leverage out of agents have built the layer around them, not just picked a vendor. A few things that actually move the cycle:
- Make tickets machine-readable. A ticket that names the files, the acceptance criteria, and the hard constraints is a spec an agent can execute. A one-line ask is a coin flip.
- Standardize context. An
AGENTS.md, a clear repo structure, and documented boundaries mean the agent starts each task from the same place instead of relearning your codebase every session. - Design review for volume. When three agents open PRs against the same module, the constraint is human judgment. Small diffs, clear ownership per feature, and a shared definition of done matter more than they used to.
- Default to fast, escalate to smart. Use a cheap, quick model for the routine 90% and reserve the expensive reasoning model for the tasks that earn it.
None of that is glamorous. All of it compounds. The orchestration tools launching this week are selling exactly this layer, which tells you the market has figured out where the value sits.
We're here to help founders and teams design and build digital products that are built to scale with you, not slow you down. If you're looking to build something, get in contact with us today!