R031 · Agent architecture › Multi-agent coordination

Resolve AI: multi-agent systems for AI-native production engineering

Read the original

multi-agentagent-architectureincident-responseobservabilityagent-reliabilityhuman-oversight

Bottom line

Resolve AI argues that the bottleneck for AI in production engineering is architecture, not model capability: debugging a live incident needs several kinds of expert knowledge at once, so it needs specialist agents run in parallel under an orchestrator, not one agent or one chat model with tools. The framing is useful and the failure analysis of single agents rings true. The post itself is a vendor position piece: it names the hard parts (coordination protocols, a DAG of information flow, continuous learning) but describes no mechanism for any of them, and gives no evaluation, error rates or customer numbers.

The post is by Spiros Xanthos (founder and CEO), Gabor Angeli and Bharat Khandelwal (research engineers) at Resolve AI, which sells an "AI SRE" for on-call and incident response. It was published 2025-09-26 and is based on a talk the authors gave to a Stanford graduate AI program.

Read it for: a clean vocabulary for where single agents break on operational work, and a list of the coordination problems any multi-agent incident system has to solve. Don't read it for: how to build one.

The argument

Coding is mostly solved; production is not. The post contrasts writing a payment service with an AI coding tool against chasing a latency spike by hand across Datadog, Loki, deploy history and timestamps. Coding work is local to a repo. Incident work spans many systems, and the knowledge needed to read each one lives in different heads.

AI-assisted vs AI-native. In an AI-assisted workflow the engineer still drives: engineer, then systems and tools, then correlation, then action, with AI speeding up single steps. In an AI-native workflow the engineer states a goal in natural language and an AI system does the correlation and proposes or takes the action. The engineer's job moves up to architecture and system design.

"Irreducible interdependence." The core claim: no single model or tool can hold expert knowledge of every relevant domain (tracing, databases, deploys, infrastructure, security, customer impact) while running a live investigation. So the work has to be split across specialists and recombined.

The maturity ladder. The post's table walks four steps and says where each breaks:

ApproachWhat it isWhere it breaks, per the post
LLMOne-shot generationSingle-pass answers, prone to hallucination, no feedback loop
LLM + toolsModel calls toolsLimited context window, no persistent state across tool calls
Single agentStateful loop with toolsInvestigates one hypothesis at a time; gets stuck on a wrong one with nothing to check it
Multi-agentOrchestrator plus domain specialistsNeeds investment in coordination protocols, communication schemas and conflict resolution

The functional case for the last row is parallel hypothesis testing: several specialists chase different explanations at once instead of one agent walking them in sequence. The post also says off-the-shelf frameworks such as LangChain don't solve this on their own, and that building it needs both domain knowledge and AI-systems expertise.

Architecture as described

The post's design is "orchestration plus individual domain specialization". For an API latency spike that hits 10x during an incident, it lists six specialist jobs that run side by side, each output feeding others through what it calls "directed acyclic graphs of information flow". Coordination protocols must prevent race conditions, deadlocks and "circular investigations", and keep "coherent narrative threads across parallel execution paths".

One goal fans out to parallel specialists, then back to one answernatural-language goalfan out, run in parallelfindingsEngineer states a goalOrchestratorassigns hypotheses, tracks the information DAGSpecialist agents (the post's latency-incident example)Tracescorrelation across 50+ servicesDatabaseslow queries, connection poolsChangesdeploys and infrastructureSecurityauth-log scanningCapacityauto-scaling behaviourCustomer impactsupport tickets, SLA contextMerged narrative, response or actionback to the engineerconflict resolution and verification: not described
multi-agent incident investigation as the post describes it

Everything above the bottom box is stated in the post; the bottom box's last line marks the step it leaves out. The six specialists are illustrations for one incident type, not a fixed role list, and the 10x and 50+ figures are illustrative, not measured.

What's missing and where it breaks

The post names the problems that make multi-agent systems hard and then stops. For an engineer evaluating this design (or Resolve's product), these are the open questions.

Failure modes to expect in a system built this way: specialists agreeing on a wrong story because each reads the others' output as fact; investigations that loop or fan out without a stop rule (the post itself names "circular investigations"); cost growing with the number of specialists on every incident, relevant or not; and actions taken on a narrative no human can audit afterwards.

Review checklist for a multi-agent incident system

Sources: The role of multi agent systems in making software engineers AI-native (Resolve AI blog, 2025-09-26).

Connections

See R031 on the knowledge map

Part of the Research notes on nowucca.com. Figures and claims are the source's own unless the summary says otherwise.