sre agent: Two Launches in One Day, and What They Mean

The sre agent category stopped being a pitch deck slide this week. On November 3, 2026, two of the biggest names in operations software shipped one on the same day: Honeycomb announced Personal and Team Agents built on its observability data, and PagerDuty announced Paige, a generally available sre agent orchestrating its newly renamed PD Reliability Platform. Both companies are making the same argument, that AI is writing code faster than humans can review it and production is where the truth comes out, so an agent that watches production is the natural response. This post walks through what each launch actually includes, where the two products differ, the caveats in the vendor evidence, and what reliability teams should take from the timing. Source

What Honeycomb shipped

Honeycomb's announcement splits the sre agent idea into two products. The Personal Agent is a collaborator for an individual engineer: it answers questions and runs investigations inside Honeycomb, keeps chats private, shows how it reached each conclusion, and acts on recommendations only when the engineer says so. Team Agents are the more ambitious half. A Team Agent is given a goal, such as watching a release, improving instrumentation, investigating alerts, improving agent quality, or maintaining knowledge docs, and works on it in the background on a schedule, a trigger, or a service level objective. Goals can be pre-defined by Honeycomb's own observability experts or written by the customer. Source

Two design choices stand out in Honeycomb's description. First, every finding links back to the query, trace, deploy, or ticket behind it, and each memory is marked as confirmed or a hypothesis. Second, the agents carry what they learn from one investigation into the next, whether the work was done by a person, a Personal Agent, or an automated run, so the team's knowledge compounds instead of living in one engineer's head. Autonomy is deliberately graduated: within the permissions a team grants, the agent can file the ticket, page the responder, or draft the fix, and Honeycomb's own framing is that autonomy grows as the agent earns it. The agents draw context from Honeycomb's high-cardinality telemetry and Model Context Protocol connectors, and both products are in early access rather than general availability. Source

What PagerDuty shipped

PagerDuty's launch is broader and immediately commercial. Paige, its sre agent, is generally available as of November 3, 2026, and works as a member of operations, engineering, and DevOps teams to detect, investigate, and remediate issues. Teams choose the autonomy level per scenario: Paige can run with a human in the loop, accelerating triage while remediation waits for approval, or act autonomously in well-understood scenarios, resolving lower-priority incidents on its own with full visibility for human operators. Like Honeycomb's agents, Paige builds a shared memory: every resolved incident feeds future performance, which PagerDuty says makes each subsequent incident easier to resolve. Source

Paige sits inside the PD Reliability Platform, the new name for what was PagerDuty Operations Cloud, which unifies signal intelligence, incident management, and automation for blended teams of agents and humans. A new console gives incident commanders a single view of incidents, on-call rotations, and which agents are enabled and what they are doing. The platform is Slack and Microsoft Teams native: when an incident triggers, PagerDuty spins up a dedicated channel and conference bridge, and once the incident resolves, a post-incident review is generated with follow-up actions pre-populated with owners and due dates. PagerDuty also launched PD Labs, an innovation center where design partners try in-progress capabilities against real problems, and released an SRE Workbench on that incubation track. Source

Pricing shifts with the platform. The PD Reliability Platform carries a platform fee based on operational scale that bundles ingested events for signal intelligence, actions for AI and automation, and seats for responders, managers, and stakeholders, replacing separate point solutions with one purchasing experience. Customers can add flexible credits and split spend between people and AI-driven actions rather than seats alone, and customers who buy an Essential, Plus, or Ultimate plan before January 15, 2027 get additional AI Actions capacity at no extra cost for the first year. Source

How the two launches differ

The philosophical gap is small but real. Honeycomb leads with telemetry depth: its agents are grounded in high-cardinality traces and MCP connectors, and the pitch is that an agent is only as good as the data it queries, so the sre agent belongs inside the observability platform that already holds the production evidence. PagerDuty leads with the incident lifecycle: Paige orchestrates detection, triage, resolution, and prevention across more than 750 integrations, and the pitch is that response coordination, not just diagnosis, is where the agent earns its keep. Honeycomb's framing also stretches beyond incidents into continuous improvement work, deploy watching, instrumentation hygiene, and reliability debt, while PagerDuty's anchors on stopping disruptions before they reach customers. Both products, notably, converged on the same two ideas: graduated autonomy that the team controls, and a shared memory that makes the agent smarter with every incident or investigation. Source

Why both vendors say now

Both announcements blame the same pressure. Honeycomb cites its own forthcoming survey in which 76% of respondents said AI-generated code has grown faster than their team's ability to review and monitor it, and 65% said they had a production issue in the past six months caused by AI rate limits, runaway agent loops, or token spend spikes in AI-built code. Treat those numbers as the vendor's own survey, self-reported and not yet published in full. PagerDuty makes the same case from its scale: it approaches one billion incidents managed annually and counts two-thirds of the Fortune 100 as customers, and it cites research that 95% of executives call reducing incidents a competitive advantage, with 77% planning to raise resilience investment in the next 12 months. The shared thesis is that AI-written code fails differently, with model drift, token runaway, and hallucination joining the usual causes, so keeping systems reliable now requires AI on the defense too. Source

The caveats in the evidence

These are launch announcements, and the supporting numbers deserve the same scrutiny we would give any vendor claim. Honeycomb's agents are early access, so the goal-oriented workflow and shared memory are described capabilities, not yet widely field-tested ones. PagerDuty points to a custom, third-party benchmark in which Paige scored among the highest at diagnosing simulated IT incidents, but the benchmark is commissioned and framed by the vendor, the scenarios are simulated rather than production incidents, and we did not find an independent replication of it. Neither announcement includes customer-controlled accuracy measurements for investigation quality, and neither says much about failure modes, such as what a Team Agent does when its hypothesis markers pile up unconfirmed, or how Paige behaves when an incident falls outside its well-understood scenarios. Early adopters should treat both products as strong drafts that need evaluation against their own incident history, not as finished judgment. Source

What this means for your team, in our assessment

This section is our read, separate from the sources. When two platform vendors ship the same category on the same day, the category is real, but the launches change the buying question more than the technical one. The question is no longer whether an sre agent is possible; it is which data foundation you want it standing on, the observability store or the incident graph, and how you will govern its growing autonomy. That governance work is exactly what we have covered before: our AI agent observability checklist covers the production signals you need before any agent can investigate well, our AI agent regression testing post covers the gates to run before you widen what an agent may do on its own, and our rogue AI agents piece covers the AgentOps controls for agents acting in production. If you adopt either product, start it in the lowest-autonomy mode, grade its investigations against your own past incidents for a few weeks, and only then expand its permissions. The vendors have made the agent easy to switch on; earning trust in it is still your job. Source