Works with the AI Agents you already use
Agent Reliability: A Production-Grade Run Loop
An agent loop is a distributed system with a non-deterministic component in the middle.
$12
Agent Reliability: A Production-Grade Run Loop
Example session with this skill installed
Here is our agent loop (a while True around messages.create that runs each tool and appends one tool_result message per call). It sometimes runs forever, sometimes refunds a customer twice after a crash, and it treats truncated answers as final. Rewrite it production-grade in loop.py.
- Read your context and instructions
- Compiled the agent reliability
loop.py: Budget(max_steps=25, max_output_tokens, max_wall_seconds) checked each iteration with distinct statuses; explicit branches for end_turn, tool_use, pause_turn, refusal, max_tokens (status 'truncated', never executed) and unknown; transient errors retried with jittered backoff, permanent ones fail fast; tool calls with timeouts and exceptions turned into is_error results, all returned in one user message; the refund tool runs sequentially, its result is checkpointed atomically by tool_use id and sent with an idempotency key run_id:tool_use_id, so a resumed run never refunds twice.
Connects securely to your tools. The creator never sees your data.
About this skill
An agent loop is a distributed system with a non-deterministic component in the middle. It needs the same engineering as any production job (bounded work, retries, timeouts, idempotency, durable state) plus LLM-specific failure modes. This skill gives your agent both, and a tested reference loop to adapt.
What your agent does with it
- Every stop reason handled explicitly: end_turn, tool_use, pause_turn, max_tokens (never success), refusal, and unknown values that stop the run instead of guessing.
- Budgets in code: steps, output tokens or money, wall-clock, tool calls per step, tool result size, each with a distinct status so dashboards can tell "finished" from "ran out".
- Retries that classify first: backoff with jitter for rate limits, overload and 5xx; fail fast on 4xx; circuit breakers for shared dependencies; SDK retries accounted for.
- Tool execution: per-tool timeouts, exceptions returned as error results, reads in parallel, writes in order, all results of a turn in one message, bounded output.
- Idempotency and resume: atomic checkpoints after every step, write results recorded by tool-use id so a crash never refunds a customer twice, idempotency keys for external writes.
- Output validation, fallbacks and concurrency limits, plus a review checklist and failure-injection test list.
Included: SKILL.md, agent_loop.py (about 350 lines: provider-neutral core with an optional Anthropic adapter), and an eval case with a naive loop to harden.
Part of the Production AI Agents series. Pair with Agent Guardrails (policy hook) and Agent Observability (events); start free with Agent Architecture.
Python reference implementation; port the structure to other languages. Provider-specific parameters in the adapter (effort, caching, fallbacks) change over time; check your SDK version. Does not decide which actions are allowed (see Agent Guardrails).
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 1 day ago
Requires
Permissions
Allowed Hosts
File Scopes
The core loop has no network dependency; only the optional Anthropic adapter calls the API, with your own credentials.