Where the work breaks.
A model can produce a useful answer and still be unsafe as an operator. Production work needs permissions, evidence, stop conditions, review gates, and a way to reconstruct what happened.
SERVICE.02 — AI AGENTS
Give an agent a job, boundaries, and a record of every action.
We build agents as controlled operating loops, not open-ended chat interfaces with access to everything.
Start with the operating problem, not the software category.
The right build becomes clear after the records, handoffs, decisions, and failure costs are visible.
What the system changes.
A bounded agent handles a defined task, uses only approved tools, records consequential steps, and stops when evidence or authority runs out.
The components serve one operating route.
Each interface and integration has a defined role inside the workflow. The system stays understandable as it expands.
Tool and role boundaries
Limit each agent to the systems, actions, data, and budgets needed for its defined job.
Human approval gates
Keep sensitive actions in draft or review mode until an authorized person approves them.
Traces and evaluations
Record inputs, tool calls, outputs, and decisions, then test important behavior against maintained cases.
Fallback and refusal behavior
Reject malformed output, pause on weak evidence, and use explicit recovery paths when a dependency fails.
One controlled slice before wider automation.
The first release proves the full route. New cases and permissions come after the current boundary is understood.
- 01 — JOB
Define one useful job.
Set the trigger, allowed inputs, expected result, and the exact point where the agent stops.
- 02 — HARNESS
Build the control layer.
Add scoped tools, schemas, policies, traces, review states, and failure handling around the model.
- 03 — EVALUATE
Earn wider autonomy.
Run real cases in observation or draft mode before expanding permissions or actions.
Inspect the closest operating evidence.
These pages show delivered systems or documented engineering patterns without exposing protected client rules or data.
Engineering approach
Inspect the harness behind a bounded agent.
See how Recode Asia connects policies, tools, traces, evaluations, and incident controls around model behavior.
Open the evidence Representative patternSee a controlled agent architecture.
The Work page shows how permissions, review points, and observable state fit around an automated loop.
Open the evidenceSet the boundary before the build.
These are the useful questions to settle before architecture or implementation begins.
What makes an AI agent different from a chatbot?
An agent can use tools and change system state. That ability requires stricter permissions, traces, review, and stop conditions.
Can the agent start in review mode?
Yes. Draft or observation mode lets the team compare proposed actions with operator decisions before granting wider authority.
Can the system refuse to act?
Yes. Low confidence, missing evidence, policy conflicts, and unavailable tools can all lead to a controlled stop or human review.
NEXT — YOUR OPERATING ROUTE
Bring one workflow that needs a better control layer.
Describe the trigger, the people involved, the current handoffs, and the result that must become dependable.