← Projects·AI Engineer
toolcall-agent
A tool-calling agent built around the ReAct loop with explicit retry/backoff, a loop guard, OpenTelemetry GenAI semantic conventions, and a fault-injection harness that forces the agent to recover from broken tool outputs. Demonstrates production reliability patterns, not prompt-demo magic.
100%
Tool / arg correctness
6 / 6
Injected faults recovered
0
Loops without bound
Why this exists
Most agent demos show happy paths. This one breaks the path on purpose. Inject a tool that returns malformed JSON for two consecutive turns; inject a 500 from the upstream API; inject an infinite-loop attempt; the agent must recover and complete the task.
What it proves
- Structured-output contracts — every tool call goes through a JSON validator before dispatch.
- Loop guard —
max_stepscap + explicit “you have already called this tool twice this turn; do something else” prompt-correction. - OTel GenAI traces — every LLM call and tool dispatch writes
gen_ai.*attributes per the CNCF semantic conventions, ready for vendor-neutral observability. - Recovery patterns — 6/6 injected faults recovered without re-prompting from a human.
What’s transferable
The loop guard, the contract validator, and the OTel trace shape apply directly to a Claude Agent SDK / MCP-backed production agent — same shape, different tool providers.