Senior Software Engineer II - Agentic Intelligence

AI agents at Honeycomb investigate, reason, and act on real observability data. They live in Canvas: the agentic workspace where engineers go to understand their systems.

The Agentic Intelligence team has shipped Canvas, the Honeycomb MCP server, and the Canvas Agent and Canvas Skills surfaces. What we're looking for today is someone who brings deep agent expertise and uses it to expand what the team can build: new agents, new surface area in Canvas, memory, spatial awareness, improved performance on our Bedrock loop.

Honeycomb's data store is fast and accepts high cardinality data; that's what makes agents built on top of it different from anything built on a conventional observability backend. This role is about taking advantage of that building agents that can do things no other observability product can do because the underlying data makes it possible.

Some of this work will start as a prototype. The expectation is that the code makes it through the full arc, from the rough first version through to something that holds up in production.

• Design and deliver production-grade agents. Build agents that investigate, reason, and act on live observability data inside Canvas. These agents must be trustworthy to engineers in high pressure situations, including mid-incident. Take one from rough first version to something that holds up under production traffic.
• Own the agent work; support the whole product. Scope, build, ship, and maintain the agents including the evals that tell you whether they got better or are just different. This role is agent-focused and also includes some fullstack development.
• Build agents only Honeycomb can build. Use a data store that returns high-cardinality queries in seconds to reason over signal a conventional backend can't serve at this fidelity correlating across services, drilling into a single trace, comparing before and after a deploy.
• Extend the surface, and decide what's next. Ship new capability into Canvas, the MCP server, and Canvas Skills memory, spatial awareness, a faster Bedrock loop and make the case for what comes after with working code. Distinguish hype from signal in a field with plenty of both.
• Define what "good" means for agents here. Set the bar: measurable against real evals, maintainable, and honest about their limits.

• Multiple agents collaborating on one shared Canvas investigation each claiming a hypothesis, publishing findings, and narrowing the search space for the others so it resolves faster (blog).
• Auto-investigation the moment an SLO burn alert fires the agent forms hypotheses and prepares visualizations before a human looks, cutting mean-time-to-insight for on-call (o11ycon 2026).
• Skills that encode a team's domain expertise e.g. Kubernetes thresholds so agents and human colleagues can lean on them (o11ycon 2026).

• AI and agent engineering experience. You've shipped LLM-based systems people relied on in production not demos, not fine-tuned models in a research context. You know where agent systems break and how to design around it.
• End-to-end ownership. On a small team there's no handoff queue. You can take something from rough prototype to production-grade without needing someone behind you to do the durable engineering.
• Current judgment, not just past experience. You have informed opinions about what's shifted in agent design in the last six to twelve months that would change how you'd build today.
• Agent architecture depth. You understand how a fast, high-cardinality data store changes what an agent can reason about, and how to design for that.
• Product judgment. You can look at what the agent layer does today and see what it should do next and make that case with a prototype, not a deck.

Observability or developer-tools background. Engineers are your users; you'll ramp faster with fluency in that world, and the work is better.

Familiarity with eval frameworks, agent tooling, RAG, and prompt engineering.