Your next colleague is an agent. Have you designed for it?
Three announcements in the past month point to the same shift. AI agents are no longer features inside a chat window. They are becoming workers in your systems, with their own identities, their own budgets and their own audit trails.
- On 8 October, Google introduced its Gemini agent for work. Each standing agent gets its own identity, its own Gmail address, calendar and file storage, and every action is logged under the agent’s name rather than a user’s. Admins can set a hard spending cap per project, and the agent pauses when it hits the limit. It can also route simpler work to cheaper models, including Anthropic’s Claude.
- In late September, Microsoft’s new Copilot moved agentic work to usage-based billing in Copilot Credits, separate from the per-user licence.
- Google’s Gemini Enterprise preview now queries BigQuery, Cloud SQL, AlloyDB and Spanner in place over MCP, using each user’s own credentials instead of copying data out.
For architects, this changes the questions we ask. Here are four we think every team should be able to answer before agents go into production.
1. Who is the agent?
If an agent acts under a human’s login, your audit trail is already wrong. Treat agents as non-human identities, with their own accounts, least-privilege roles and a named human owner. Your identity provider, not the prompt, is where access is decided.
2. What is it allowed to spend?
Per-seat licensing hid AI cost inside a flat fee. Usage billing does not. Budgets per project, alerts, hard caps and chargeback to the team that owns the agent should be designed in from day one, the same way we learned to do FinOps for cloud.
3. Where does the data stay?
Querying data where it lives, under the caller’s own permissions, is a better default than copying it into a vector store with flattened access. Protocols such as MCP make this practical. Ask whether your agent design respects the row and column security you already have.
4. Which model does each step need?
Routing is now an architecture decision. A planning step may need a frontier model, while classification or extraction can run on a small, cheap one. A routing layer you control also protects you when a vendor retires a model, which is happening more often (OpenAI’s GPT-4o and GPT-4.1 leave its API on 14 January 2027).
The takeaway
The agent platforms are maturing quickly. The organisations that benefit most will be the ones that bring the same discipline they use for people and services (identity, budgets, data boundaries and observability) to their agents.
At KSL, we help teams design agent architectures that are secure, auditable and cost-aware from the first release. If you are moving from pilots to production, we would be glad to compare notes.