Time since we first saw it
Day 0
Early. Similar jobs usually come down between day 2 and day 14.
From 955 similar postings that came down. Coming down is not always a hire.
Employment type in the employer’s posting: Full timeFrom the employer’s posting
Build directly against frontier LLM SDKs and agent layers: Anthropic's Claude SDK and Agent SDK, OpenAI's Agents SDK and Responses API, and Google's Gemini/GenAI SDK.
Design agent capability surfaces: tool registries, agent registries, a lazy-loading skill system, and versioned skill promotion through immutable artifacts.
Build retrieval and grounding: document parsing, metadata-first ingestion, vector search, and end-to-end citation pipelines for documents and web results.
Build and extend multi-agent orchestration on Microsoft Agent Framework, LangGraph, and Deep Agents, including stateful graphs, checkpointing, sub-agents, human-in-the-loop interrupts, and long-running or background execution. Build directly against frontier LLM SDKs and agent layers: Anthropic's Claude SDK and Agent SDK, OpenAI's Agents SDK and Responses API, and Google's Gemini/GenAI SDK. Own the model-agnostic layer: routing across providers, adaptive routing logic, extended-thinking support, cost and token accounting, and graceful degradation as models and SDKs change.
Own the model-agnostic layer: routing across providers, adaptive routing logic, extended-thinking support, cost and token accounting, and graceful degradation as models and SDKs change. Design agent capability surfaces: tool registries, agent registries, a lazy-loading skill system, and versioned skill promotion through immutable artifacts. Build retrieval and grounding: document parsing, metadata-first ingestion, vector search, and end-to-end citation pipelines for documents and web results.
Design agent capability surfaces: tool registries, agent registries, a lazy-loading skill system, and versioned skill promotion through immutable artifacts. Build retrieval and grounding: document parsing, metadata-first ingestion, vector search, and end-to-end citation pipelines for documents and web results. Build evaluation infrastructure: offline eval suites, regression gates on prompt and model changes, and tracing for diagnosable agent failures.
Chubb is the world's largest publicly traded property and casualty insurer.
In the employer’s words · Read in context
Actively hiringFirst seen by us today, still listed, and not a repost of an earlier posting.
These dates show when we found the posting. Check the employer’s website to confirm it is still accepting applications.
Day 0
Early. Similar jobs usually come down between day 2 and day 14.
From 955 similar postings that came down. Coming down is not always a hire.
33 open jobs
Open jobs by the week we first saw them, last 12 weeks.
Hiring recordAdd your resume and we’ll mark which of these tools you already use, and point out anything in the posting that could rule you out.
- 7+ years building production software, with recent, substantial work on LLM-based systems. - Strong TypeScript and working proficiency in Python (FastAPI, async, Pydantic). - Demonstrated production depth in Microsoft Agent Framework, LangGraph, or Deep Agents.
- Keep clean separation between business logic and infrastructure; maintain type safety, error handling, idempotency, and environment-agnostic configuration. - Build for operability with structured logging, OpenTelemetry tracing, Azure Application Insights, meaningful SLOs, and instrumentation for SSE streaming failures and memory-growth patterns. - Design multi-tenant systems with hierarchical entitlements and role models, per-user OAuth delegation (MSAL/PKCE), and rate and quota enforcement.
- Fluency in software design patterns. - Cloud-native delivery: containers, Kubernetes (AKS preferred), CI/CD, and production observability. - Ability to work through ambiguity with stakeholders and turn a vague capability ask into a defensible technical design.
- Sandboxing and process isolation: OS security primitives, gVisor, nsjail, Firecracker, Seatbelt, Bubblewrap, or similar. - Desktop application engineering, retrieval systems, vector indexing, citation-grade grounding, eval and LLM observability tooling, React, Nx monorepos, PostgreSQL/Cosmos DB, Redis, regulated environment experience, and insurance or financial services domain exposure. How you work
Complete your application on fa-ewgu-saasfaprod1.fa.ocs.oraclecloud.com. The employer’s form will show what is required.
Already applied? Track this application
ABOUT CHUBB
Chubb is the world's largest publicly traded property and casualty insurer. With operations in 54 countries and territories, Chubb provides commercial and personal property and casualty insurance, personal accident and supplemental health insurance, reinsurance, and life insurance to a diverse group of clients. The company is defined by its extensive product and service offerings, broad distribution capabilities, exceptional financial strength, underwriting excellence, superior claims handling expertise, and local operations globally.
Overview
Chubb AI builds the enterprise AI platform for Chubb. The platform is a multi-tenant, model-agnostic environment that spans a web application, a desktop agent client, an AI App Store, a skills and extension ecosystem, and an API/tool layer over Chubb's knowledge.
The role
Design and build the agent infrastructure at the core of the platform: orchestration engines, the desktop agent harness, the skills and tool registries, and the policy layers that make it safe to run inside Chubb. This is a hands-on engineering role with architectural ownership, working closely with platform engineering, DevOps, security, and Model Risk Governance.
What you'll work on
Build and extend multi-agent orchestration on Microsoft Agent Framework, LangGraph, and Deep Agents, including stateful graphs, checkpointing, sub-agents, human-in-the-loop interrupts, and long-running or background execution.
Build directly against frontier LLM SDKs and agent layers: Anthropic's Claude SDK and Agent SDK, OpenAI's Agents SDK and Responses API, and Google's Gemini/GenAI SDK.
Own the model-agnostic layer: routing across providers, adaptive routing logic, extended-thinking support, cost and token accounting, and graceful degradation as models and SDKs change.
Design agent capability surfaces: tool registries, agent registries, a lazy-loading skill system, and versioned skill promotion through immutable artifacts.
Build retrieval and grounding: document parsing, metadata-first ingestion, vector search, and end-to-end citation pipelines for documents and web results.
Build evaluation infrastructure: offline eval suites, regression gates on prompt and model changes, and tracing for diagnosable agent failures.
Engineer a meta-harness that runs multiple underlying agent runtimes behind one uniform API and event stream.
Build adapter layers for harnesses, normalize SDKs/event models/tool schemas/hook systems/session and permission semantics, and support a runner/server split over WebSocket.
Build MCP integration depth, including stdio and remote SSE servers, first-party servers such as Outlook and Atlassian, and a tool/UI bridge for sandboxed extension widgets with a manifest and permission model.
Implement the permission system with stateful ALLOW / ASK / DENY decisions driven by policy through hooks.
Work on sandboxing and process isolation, including restricted tokens, namespaces, Seatbelt/Bubblewrap-class mechanisms, container and microVM options, filesystem policy design, per-session ACLs, and platform-specific drivers.
Build scheduling, memory, and continuity features such as recurring tasks, cross-conversation reference, file indexing and mentions, and a runtime-agnostic persistent memory contract.
Apply structural patterns deliberately: facade, adapters, ports-and-adapters, factories, dependency injection, and registries.
Keep clean separation between business logic and infrastructure; maintain type safety, error handling, idempotency, and environment-agnostic configuration.
Build for operability with structured logging, OpenTelemetry tracing, Azure Application Insights, meaningful SLOs, and instrumentation for SSE streaming failures and memory-growth patterns.
Design multi-tenant systems with hierarchical entitlements and role models, per-user OAuth delegation (MSAL/PKCE), and rate and quota enforcement.
Write the design docs.
What we're looking for
7+ years building production software, with recent, substantial work on LLM-based systems.
Strong TypeScript and working proficiency in Python (FastAPI, async, Pydantic).
Demonstrated production depth in Microsoft Agent Framework, LangGraph, or Deep Agents.
Direct experience with at least one frontier provider, SDK, at the agent-loop level.
Hands-on experience with tool calling, MCP or a comparable tool protocol, streaming (SSE/WebSocket), and structured output.
Fluency in software design patterns.
Cloud-native delivery: containers, Kubernetes (AKS preferred), CI/CD, and production observability.
Ability to work through ambiguity with stakeholders and turn a vague capability ask into a defensible technical design.
Good to have
Google ADK, Mastra, Pydantic AI, Semantic Kernel, CrewAI, LlamaIndex, AutoGen.
Other frontier LLM SDKs and agent layers as they emerge, including open-weight and non-US providers.
LiteLLM or comparable multi-provider gateways and routers.
Building on or extending Pi Coding Agent, Claude Code, OpenCode, Kimi Code, Codex or similar, especially embedded via SDK.
Sandboxing and process isolation: OS security primitives, gVisor, nsjail, Firecracker, Seatbelt, Bubblewrap, or similar.
Desktop application engineering, retrieval systems, vector indexing, citation-grade grounding, eval and LLM observability tooling, React, Nx monorepos, PostgreSQL/Cosmos DB, Redis, regulated environment experience, and insurance or financial services domain exposure.
How you work
You design with standard software engineering patterns before you code.
You research before you commit, and you write down what you found.
You prefer production-ready over clever.
You give and take direct technical feedback well.
You care that the thing works for the person using it.
What we offer in the team
Ownership of platform foundations used across a global organization.
A team that treats architecture as a first-class activity.
Access to frontier models and tooling, and a mandate to figure out what they're good for.
Chubb Canada does not use artificial intelligence (AI) tools to assess, screen, or select applicants.
At Chubb we are committed to providing equal employment opportunities to all employees and applicants. It is our policy to provide equal employment opportunities to employees and applicants based on job-related qualifications and ability to perform a job. If you require an accommodation during the hiring process or upon hire, please inform Human Resources. If a selected applicant requests accommodation during the recruitment process, Chubb will consult with the applicant in order to provide suitable accommodation that takes into account the applicant’s accessibility needs.
We file this job under AI Engineer, by the work the posting describes.
AI engineers build products on top of AI models that already exist, most often large language models: they connect a model to a company's data and tools and ship the assistants, agents and search features people use.
Postings ask for strong Python or TypeScript, experience with model APIs, retrieval and agent frameworks, and ways to test whether the model's answers are good enough.
An ML engineer trains and serves models; an AI engineer usually starts from a model someone else trained and builds the product around it, though titles often mix the two.
Selected passages from the saved posting. Check the full description for conditions and exceptions.
No pay amount identified in the saved description.
Canada
- Build adapter layers for harnesses, normalize SDKs/event models/tool schemas/hook systems/session and permission semantics, and support a runner/server split over WebSocket. - Build MCP integration depth, including stdio and remote SSE servers, first-party servers such as Outlook and Atlassian, and a tool/UI bridge for sandboxed extension widgets with a manifest and permission model. - Implement the permission system with stateful ALLOW / ASK / DENY decisions driven by policy through hooks.
No clear work-authorization passage found. Eligibility is unconfirmed.