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:
| Approach | What it is | Where it breaks, per the post |
|---|---|---|
| LLM | One-shot generation | Single-pass answers, prone to hallucination, no feedback loop |
| LLM + tools | Model calls tools | Limited context window, no persistent state across tool calls |
| Single agent | Stateful loop with tools | Investigates one hypothesis at a time; gets stuck on a wrong one with nothing to check it |
| Multi-agent | Orchestrator plus domain specialists | Needs 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".
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.
No coordination mechanism. "Formal coordination protocols", "communication schemas" and "conflict resolution" are named as requirements, but no message format, ownership rule or merge step is described. How two specialists' contradictory findings are reconciled is the core design question, and it is not answered.
No shared-state model. Agents are "stateful" and keep "investigation context", but there is no word on whether that is a blackboard, a shared timeline, a graph of hypotheses or per-agent memory, or on how stale or wrong findings are retracted.
No verification of the system's own conclusions. The ladder criticizes single agents for having no feedback loop, yet nothing says how the multi-agent system checks a root-cause claim before an engineer acts on it. Parallel agents can produce parallel confident errors.
No baseline. The "sequential reasoning constraint" is presented as fundamental, but one agent with parallel tool calls, or best-of-N single agents at the same token budget, can also test hypotheses in parallel. Without that comparison, the case for many agents over one well-equipped agent is asserted, not shown.
No evaluation or numbers. Salesforce, Zscaler and Coinbase are named as customers with better "velocity and reliability"; no metric, incident count, time-to-resolution or error rate is given.
Human role is thin. Engineers "set goals". There are no approval gates for actions, no escalation rule and no way to see why an agent concluded what it did, which matters most during an incident.
Learning is one sentence. The system should "learn continuously both from interactions and ephemeral failure modes", with no mechanism, review step or guard against learning the wrong lesson.
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
Is there a matched-budget baseline (one agent with parallel tools, or best-of-N) that the multi-agent design beats on real incidents?
Is the information flow written down: which agent's output feeds which, and is it really acyclic?
Who owns each hypothesis, and how is a duplicate or abandoned line of inquiry detected?
How are contradictory findings resolved, and is the resolution recorded with its evidence?
Does every root-cause claim cite the signals (traces, logs, deploys) it rests on, so a human can check it in minutes?
Is there a stop rule (time, tokens, confidence) for the whole investigation and for each specialist?
Which actions can agents take on their own, which need approval, and is that enforced outside the model?
What does the system learn from an incident, where is it stored, and who reviews it before it changes behaviour?
Are time-to-resolution, false root causes and cost per incident measured over time?
Related entries and sources
R019 s3 swarms: the opposite approach, coordination learned instead of designed, and the matched-budget best-of-N baseline that this post never runs.
R021 DeepSeek's Super-agent: a concrete, hand-written coordination protocol (a router, a ticket per task, subscriptions for follow-ups), the kind of mechanism this post names but doesn't show.
R015 Workflows over agents: argues extra agents earn their keep only for parallel work, which is exactly the post's one functional argument.
R016 Human verification of AI output, and R020 the model proposes, the harness enforces: where verification of an agent's conclusions should live.
R006 harness terms: the vocabulary (agent loop, tools, context management, permissions) behind the post's maturity ladder.
Sources: The role of multi agent systems in making software engineers AI-native (Resolve AI blog, 2025-09-26).