AI agents and model integrations
Platform · AI agents
Platform concern — Build agent execution as a controlled workflow with tools, budgets, traces, and evaluation—not as an unbounded chat loop.Model adapters, permissioned tools, durable state, streaming, and measurable quality.
Key concepts
Model
Probabilistic
Tool
Deterministic effect
Evaluation
Quality signal
Operating model
Decisions to make before production
Never let model text directly become a database query, shell command, permission decision, or external message.
Boundary and failure behavior
Agent workloads combine nondeterministic model output with deterministic application side effects.Questions for review
Implementation notes
Decision lens
Related topics
- stream · Stream agent work — Read the focused guide for this boundary.
- shield-halved · Protect tools — Read the focused guide for this boundary.
- gauge-high · Benchmark agents — Read the focused guide for this boundary.