Guides
The decisions involved in building an agent inside a product.
Choose the work#
Find a recurring job, decide which tools matter, and define a useful pilot.
- What is a product agent?The model is the engine; the product is the harness: the context and objects, the acting identity and permissions, the screens where people review, and the receipts, history, and undo.4 min
- Build your first product agentA product team takes one job from historical cases to a reviewed pilot: tools, product architecture, permissions, background work, and a business scorecard.4 min
- Prioritize tools from historical casesA labeling sheet and a worked capability backlog turn support cases, operator work, and traces into decisions about which tools to build next.4 min
Build the agent#
Frameworks, product context, durable jobs, and external entry points.
- Choose an agent framework for the product jobAI SDK, OpenAI Agents SDK, LangGraph, or an existing job system: compare approval, saved state, deployment, and restart behavior using the same acceptance tests.4 min
- Background runs and durable streamsOne conversation can hold several jobs. Each job needs an ID and one owner, so a lost reply, a replaced worker, or a closed tab doesn’t turn into duplicate or invisible work.6 min
- Your product through MCPMCP moves where a request starts, not who owns the operation. The product still resolves the account, checks permission, and returns a result the host can’t misread.5 min
Set permissions#
Acting identities, approval contracts, capability ownership, and rollout.
- Authorization for product agentsThe requester, reviewer, and acting account may differ. A practical execution contract binds approval to exact work and checks current authority before the effect happens.3 min
- Skills and a shared capability platformMultiple teams can contribute product jobs through shared contracts for tools, skills, ownership, permissions, versions, review, and rollout.5 min
Design the experience#
Review cards, bulk changes, recovery, history, mobile, and proactive work.
- Approval cards and bulk reviewAn approval should be tied to one exact proposal: the items, the changes, and the account that acts. The product also has to decide what happens when the world changes after the review.6 min
- Undo and partial recoveryUndo is a new operation. Recovery should restore only what recovery safely can, skip records someone edited since, and say what can’t be reversed.5 min
- Agent history and receiptsA receipt from the system that made the change, not the agent’s summary, lets a second person see what happened without the original chat.4 min
- Mobile review and proactive notificationsA proactive finding becomes a saved proposal, a small push payload, an authenticated review, and a receipt. Handle delayed delivery, expired work, and device switches.4 min
- Proactive agents and standing rulesA standing instruction isn’t permanent authority. The rule names its scope, acting account, limits, and owner, and the product checks permission again at the moment of action.5 min
Operate and measure#
Inspect traces, investigate failures, and measure the whole job with BizOps.
- Agent tracing and operational reviewWhat a trace contains, four tracing tools worth trying, what to log beyond the model call, and a review routine a small team can start this week.9 min
- Measure agent impact with BizOpsFinance, product, and engineering agree on checked results, human effort, full cost, cohorts, and late corrections before deciding whether an agent pilot worked.5 min