Agent API

Many “agent quality” issues are actually connection issues:

• output quality drifts day to day
• cost spikes without clear attribution
• identical tasks behave differently across sessions

Agent API reduces that uncertainty: you define the route, the runtime behaves predictably.

The three things that matter most

• capability: which models and reasoning level you can use
• stability: latency and reliability profile
• cost: budget and quota control

When to switch connections

• quality drops and the task itself hasn’t changed
• latency jitter starts hurting delivery speed
• current pricing no longer fits the workflow

Practical setup strategy

• keep one primary connection per workspace
• separate production and experiment connections
• lock connection and default model before prompt tuning

Rule of thumb: stabilize the road first, then optimize the driving.

Under the hood

Configure in Settings → Agent API. Typical fields:

• name
• baseUrl
• apiFormat
• models / defaultModel / customModel

Credential storage:

• config: ~/.tentarc/config.json
• secrets: ~/.tentarc/credentials.enc

Model resolution order:

session model -> active connection default -> workspace defaults.model -> global config model -> DEFAULT_MODEL