Plan act observe
Use a tool, inspect the result and choose the next step. Stop at the budget or when help is needed.
AGENT DESIGN
The right amount of AI.
For the work that matters.
One agent can use many tools and skills. A workflow can have several AI steps without a swarm. We choose the smallest system that meets your quality, speed and cost targets.
FIND A STARTING DESIGN
Tell us how
the work behaves.
This guide suggests a starting architecture. A real decision needs representative examples, access checks and measured comparisons.
- GraphThe map of steps, branches and handoffs.
- LoopRepeat a bounded step until the work is ready, blocked or over budget.
- TeamSeparate agents coordinate distinct work. Add them for a reason.
Five designs. Different jobs.
| Design | Good starting point when | Business example |
|---|---|---|
| Rules and software | Known decisions and exact calculations. | Copy approved records, calculate totals, enforce limits. |
| Workflow + bounded AI | Known route; variable input. | Extract an invoice, check the math, route exceptions. |
| One agent + tools | One adaptive case with shared context. | Investigate a missing order and prepare a response. |
| Agent + verifier | A concrete output benefits from an independent check. | Review a proposal’s evidence and calculations. |
| Coordinated team | Independent work or distinct context has measured value. | Research acquisition workstreams and reconcile findings. |
Department ownership does not require a separate agent for every department. Context separation does not establish security isolation; permissions must be enforced in the tools, storage and runtime.
Example execution graph. Missing information pauses the case for clarification; checks and approvals determine which step can run next. Retries and revision loops have explicit limits.
LOOPS WITH LIMITS
Check the work.
Keep people in control.
Generate check repair
Test an output against rules and evidence. Bound retries; repeated failure becomes an exception.
Pause review resume
Save the case, ask the right person, then recheck facts before executing an approved action.
Measure improve release
Compare against a test set and the baseline. Deploy reviewed changes with a rollback path.
DON’T PAY FOR EMPTY COMPLEXITY
When one agent
makes more sense.
Short conversations. Sequential work. One shared case history. Tight latency or cost limits. A common customer voice.
- Less context copyingAvoid repeating or summarizing the same history at every handoff.
- Fewer model startsReduce orchestration that adds latency without useful parallel work.
- One controlled write pathAvoid conflicting updates and duplicate actions.
- Easier improvementFewer prompts and interfaces can be easier to test, maintain and upgrade.
OUR ACCEPTANCE TEST
Count useful results.
Count all the costs.
Compare accepted completion, review effort, typical and slow-case latency, incorrect actions, source support and total cost per accepted result. Include failed attempts and rework. Expand a team only when it earns its place.
BUILD YOUR WORKSPACE
Bring us the work.
We’ll build the system.
Start with one useful workflow. Prove the result. Expand what works.
Free 30-minute call with Oliver. Scope, running costs and project price agreed before work starts.