Gemini 3.5 Flash GA and Managed Agents preview: what changes for a developer this week

What it is

Two things landed on the Gemini API changelog that matter if you build agents. Gemini 3.5 Flash is Google's fast tier model, now documented on the Gemini API developer site, and Managed Agents is a new hosted agent orchestration layer in the Gemini API, currently in public preview. Flash is a model swap. Managed Agents is a bet that Google can run your tool-calling loop for you.

The next-work-session test

You have a retrieval + tool-use agent glued together with your own scheduler, a queue, and a JSON-mode fallback for when the model forgets the schema. Your next session:

What changes: if the managed loop matches your hand-rolled one on your evals, you throw away a scheduler. If it doesn't, you learned that in one afternoon.

Pricing

Gemini 3.5 Flash pricing is not on a single clean Google page I could verify end-to-end for the AI-for-developers tier — the Agent Platform pricing page is the authoritative Google surface, but it covers the Enterprise Agent Platform SKUs. Third-party trackers list Flash at roughly $1.50 input / $9.00 output per million tokens, and one commentary piece frames 3.5 Flash as costing about three times the model it replaced. Treat those numbers as third-party claims until you see them on your own billing console.

Managed Agents pricing during preview: unverified. Google has not published a separate per-agent-invocation rate that I could confirm; assume you pay for the underlying model tokens plus any tool-call overhead, and check your billing before running a load test.

What we'd actually use it for

Two narrow things:

• Flash as the cheap step in a two-model pipeline. Router / classifier / "does this need Pro?" gating. Not as the reasoning model. The price bump from the previous Flash makes "use Flash for everything" a worse default than it was six months ago.
• Managed Agents for internal tools where you own both ends. A single-purpose agent with 3–6 tools and a clear success metric. Not your customer-facing agent. Not until you've watched the preview for a quarter.

Limits

• Public preview means the API surface can change. Don't build a paying customer feature on it this month.
• You still write the tools and the evals. The managed loop doesn't remove the two hardest parts of shipping an agent.
• Hosted orchestration is a lock-in vector. Every agent you rebuild against Managed Agents is one you'll port later if you leave.
• Debugging a loop you don't run yourself is harder. Check what tracing and step-level logs are actually exposed in the agents overview docs before you commit.
• Pricing opacity during preview. If you can't model cost per successful task, you can't compare it to your current stack.

Try it if

• You already ship on the Gemini API and have an eval set you trust.
• You have one small internal agent whose failure mode is embarrassing but not catastrophic.
• You're spending real engineering time on the orchestration layer itself, not on the tools or the prompts.
• You want a cheap A/B against your current Flash-tier calls this week.

Skip it if

• Your agent stack is already boring and works. Managed Agents in preview is not worth the churn.
• You need portability across model vendors. Keep your own loop.
• Your bottleneck is tool quality or retrieval, not orchestration. A hosted loop won't fix that.
• You can't get finance to sign off on unclear preview pricing.

The honest read: Flash GA is a five-minute config change worth running through your evals today. Managed Agents is worth one afternoon on one narrow agent, and a note in your calendar to re-check when it exits preview.