Gemini agent vs custom AI agent: which work goes where

Since 8 October, one question keeps coming up in conversations with founders and ops leads: Google now has a universal agent for work, so do we still need a custom one?

The useful question isn't which agent is smarter. It's which work goes where. Some work belongs with a general-purpose agent that already lives inside your email, documents and chat. Other work touches systems, channels and approval rules no general agent can see. Here is how to split the two, control the cost, and pilot both without lock-in.

At Gemini at Work 2026 on 8 October, Google Cloud announced the Gemini agent, described as a "new single, universal agent for work" across the Gemini Enterprise app, Workspace and third-party services. According to 9to5Google (opens in a new tab), it answers questions, handles knowledge work, creates media, and writes and runs code.

Put simply, you give it a job to do, not a script to follow, and it breaks the job down itself. The main pieces:

• Sub-agents. Multi-step tasks are split across sub-agents that can run in parallel or in sequence, for hours or days.
• Model choice. Each job runs on the best-fit model, and today that includes both the Gemini and Claude families.
• Memory. It has four memory types: session, semantic, procedural (including skills it writes for itself) and episodic.
• Surfaces. You can call it with @Gemini in Gmail, Docs, Sheets, Slides and Chat. It also works from the command line, Slack and Microsoft 365, and an API lets it run as a headless agent inside third-party apps.
• Roles. It can work as a team member or act on behalf of a role, such as a finance analyst.

It's in private preview, and wider availability is "coming soon" for select Workspace Business and Enterprise plans. Futurum Group (opens in a new tab) notes that Google didn't disclose pricing or general availability dates for most of the announcements.

The Gemini agent is built for the collaboration layer: documents, threads, meetings and shared data. Think drafting a board deck from last quarter's numbers, finding a meeting slot across three calendars, or writing a weekly report from connected data.

SiliconANGLE (opens in a new tab) lists the connectors Google named: Confluence, Microsoft Office, Teams, Slack, Workspace, Git, Jira, Salesforce, ServiceNow, BigQuery, Databricks, Postgres, Snowflake and desktop files. If your team's work mostly lives in those tools, a universal agent may cover a lot of it.

Coworker sub-agents get their own identities (@agents.company.com email addresses) and persistent storage, and see only the context users give them. Reusable skills can be published to a company registry, and saved BigQuery reports can be re-run without extra token costs.

Thomas Kurian made the point himself, as SiliconANGLE reports: to deliver outcomes, an agent has to be connected to personal workflows, systems of record and enterprise controls. That is also the test for where a universal agent stops.

Many SMEs don't run on Snowflake and ServiceNow. They run on Tally, a Shopify store, a regional ERP, a folder of supplier PDFs and a Postgres database someone set up years ago. Of these, only Postgres appears in the named connector list.

When the work has to read from or write to those systems, you need something built for them. An invoice-to-Tally agent is a good example of work the generic connectors don't cover. So is a talk-to-your-data assistant over your own databases, where the schema, the business terms and the access rules are specific to you. This is the core of our integrations work: connecting agents to the systems of record your team already runs, not asking you to migrate.

Raising a purchase order, issuing a credit note or changing a price needs the same steps every time, plus a named person signing off before anything is sent. An agent that plans its own approach is the wrong tool for a process that must not vary.

A custom agent can run a fixed workflow where the model handles judgement, such as reading an RFQ or matching line items, and the system enforces the rest. RFQ and PO automation with human approval steps follows this pattern.

The announced surfaces are internal: Workspace, Slack, Microsoft 365, the command line and an API. Your customers aren't in your Workspace. If the work happens over WhatsApp, on your storefront, or in Hindi, Arabic or another regional language, you will need to build the channel, tone and escalation rules yourself.

Nightly reconciliations and scheduled supplier follow-ups run without anyone watching. That work needs its own schedule, failure alerts and logs your team can read the next morning.

Custom agents used to mean a lot of plumbing. The Claude Platform release notes (opens in a new tab) show several relevant releases for Claude Managed Agents over the last few months:

• Webhooks (30 June): events across the agent, deployment and deployment-run lifecycle, such as a paused deployment or a failed scheduled run, without polling.
• Dynamic workflows (9 October, beta): for work with many pieces, such as reviewing hundreds of documents, the agent writes a workflow that runs many agents in phases and combines the results in the background.
• OpenTelemetry export (14 October, beta): traces for Messages API requests and Managed Agents sessions, so they can go into the observability stack you already use. Scheduled deployments became available on Claude Platform on AWS the same day.
• Org-level web domain controls (20 October): admins can turn off web fetch or set allowed and blocked domains, and Managed Agents follow those restrictions.

For teams on Google Cloud, the release notes also cover Claude Models on Google Cloud (Anthropic managed), where Google Cloud IAM controls access and Google Cloud Marketplace handles billing.

Softpiper is a member of the Claude Partner Network · Select Services Partner, and we build custom AI agents and automation on Claude and Gemini. The takeaway: scheduling, tracing and failure alerts are now platform features, not things you build from scratch.

Agents that run for hours can burn through tokens unnoticed. Both sides now offer real controls; switch them on from day one.

On the Google side:

• Hard AI spend limits are set in the Cloud Billing Console, and the agent pauses when a cap triggers. Per-project tracking makes department chargeback possible (SiliconANGLE (opens in a new tab)).
• Smart routing sends simple tasks to Gemini Flash and long-horizon work to frontier models.
• The Gemini Enterprise app now offers pay-as-you-go alongside per-user seats, so agent workloads don't hit quota limits mid-task. Flexible Savings Plans take 10–20% off token costs for a committed monthly spend, with no minimums or maximums. Off-peak capacity for work that can wait costs up to half the inference price (Google Cloud Blog (opens in a new tab)).
• Governance includes agent identity, OAuth-propagated permissions, per-agent audit trails, Agent Sandbox and Agent Gateway (Futurum Group (opens in a new tab)).

On the custom side: domain allow-lists, OpenTelemetry traces and lifecycle webhooks give you the same basics: what the agent can reach, what it did, and when it failed.

Seats suit steady, everyday use across a team; pay-as-you-go suits bursty agent jobs. Either way, set the cap before the pilot starts, not after the first invoice. We set up budget caps, monitoring and deployment on Google Cloud as part of agent projects for this reason.

Delegate to a universal agent when:

• The work lives in Workspace, Microsoft 365, Slack or the named connectors.
• The output is a draft a person will review: a deck, a summary, a schedule, a report.
• Variation between runs is acceptable, or even useful.

Build a custom agent when:

• It must read from or write to a system outside the connector list (Tally, Shopify, a regional ERP, your own database).
• The steps must be the same every time, with a named human approval before anything leaves the building.
• It faces customers, runs over WhatsApp, or works in a regional language.
• It runs unattended on a schedule and needs its own alerts and traces.
• You need it now, not when a private preview reaches your plan.

Combine them when: the request starts in the collaboration layer but the real work happens in a specialist system. The universal agent acts as the front door, and a custom agent sits behind an API as the specialist. Google hasn't yet published enough detail to say exactly how the Gemini agent will call your own services. The low-risk version today is a shared data layer: the custom agent writes clean results into Postgres or BigQuery, and the universal agent reports over them through its connectors. Your specialist logic stays portable either way.

If the lines aren't obvious for your business, we can help you decide what to delegate and what to build.

1. Pick one workflow. Choose one with a clear owner, a measurable output and real volume.
2. Set a budget cap. Use hard spend limits on the Google side and a monthly ceiling on the custom side. Decide in advance what happens when the cap is hit.
3. Instrument it. Turn on audit trails and OpenTelemetry tracing. Record time saved, error rate and how often a person had to step in.
4. Keep the logic portable. Store prompts, rules and approval steps in your own repository, and treat models as interchangeable. The Gemini agent already routes across Gemini and Claude models, and Claude models are available on Google Cloud.
5. Review after 30 days. Compare the cost per completed task with the old process. Then expand it, change it or stop it.

For an example of the kind of product work we do, see Infra-Q, an AI web product in development.

A universal work agent will probably absorb much of the everyday collaboration work in companies on Workspace or Microsoft 365. The work that sets your business apart, in your ledgers, storefront, supplier threads and customers' languages, will still need agents built for it. If you'd like to map your workflows onto both, book a call with the Softpiper team.

• Google Cloud announces 'Gemini agent' as 'universal agent for work' (9to5Google, 8 Oct 2026) (opens in a new tab)
• Google Cloud introduces Gemini agent to change enterprise work (SiliconANGLE, 8 Oct 2026) (opens in a new tab)
• Google Collapses Enterprise AI Into a Single Gemini Agent at Gemini at Work 2026 (Futurum Group) (opens in a new tab)
• Flexible billing and cost controls for agents on Google Cloud (Google Cloud Blog, 26 Aug 2026) (opens in a new tab)
• Claude Platform release notes (Anthropic) (opens in a new tab)