Sign InOpen Brain
AI EngineerVideoSource Linked

AI Agents Are Just Distributed Systems Now — Salman Munaf, TikTok

Once agents mutate external state, timeouts mean unknown outcomes. Builders need idempotent tools, bounded retries, scoped credentials, durable traces, and explicit recovery paths.

AI Engineer · Aug 29, 2026
Open Source Open MarkdownOpen JSON
Source Summary

A tool timeout does not prove failure: the remote side may already have committed the action. The talk recommends request IDs, **idempotency keys**, status lookups, circuit breakers, compensation operations, and limits on turns, spend, and parallel calls.

Practical Implication

Design every agent tool like a distributed-system boundary. Persist each step, define the source of truth, treat memory as an invalidatable cache, and bind approvals to the exact action, actor, timestamp, parameters, and expiration.

Agent-Ready Context
A tool timeout does not prove failure: the remote side may already have committed the action. The talk recommends request IDs, **idempotency keys**, status lookups, circuit breakers, compensation operations, and limits on turns, spend, and parallel calls.

Design every agent tool like a distributed-system boundary. Persist each step, define the source of truth, treat memory as an invalidatable cache, and bind approvals to the exact action, actor, timestamp, parameters, and expiration.

A stronger model can reduce reasoning mistakes but cannot remove network ambiguity, stale state, or adversarial input. Some irreversible actions cannot be truly undone, so compensation and human approval remain domain-specific safeguards.
Connected Context · Feed7 Judgment

This shifts agent reliability from better reasoning alone to explicit distributed-systems semantics at every tool boundary. Persisted steps, idempotency, status reconciliation, bounded execution, exact approvals, and compensation address ambiguity that model upgrades cannot remove; irreversible actions and stale external truth still require domain-specific controls.

Tribal Dungeons of Global Shipping: AI Agents at Global Scale — Dmitry Buykin, MaerskMaersk’s slow legacy calls and production rights provide operating-scale support for persisted traces and bounded tools where a timeout cannot safely be treated as failure.How Kepler Built Verifiable AI for Financial Services — Vinoo GaneshKepler reinforces the source-of-truth boundary by keeping numeric mutation in deterministic, replayable systems instead of model state.Healthcare’s Agent Bytecode: X12 as the Harness for AI Agents — Vasant Kearney, OnlayX12 demonstrates how typed protocols can make tool actions and lifecycle transitions checkable, while its stale payer data confirms that protocol validity does not guarantee current truth.Learning on the Job: The Future of Post-Training — Raymond Feng, Applied ComputeThe difficulty of learning from non-replayable production interactions strengthens the case for persisting each step and exposing status, while showing that traces alone do not restore missing counterfactuals.
Context Map
agentcoding#harness-engineering#tool-use#agent-reliability
Uncertainty
A stronger model can reduce reasoning mistakes but cannot remove network ambiguity, stale state, or adversarial input. Some irreversible actions cannot be truly undone, so compensation and human approval remain domain-specific safeguards.