Sim’s Agent Picks the Resumes Humans Actually Read

How Sim Filters Every Candidate

Every application that lands in Sim's inbox meets the same gatekeeper first: an AI agent built on the same workspace the company sells. The agent does not decide who gets hired. It decides who gets read.

Sim is hiring for seven roles across engineering, design, and go-to-market (five listed publicly as of August 2026, two held closer to the vest), and every candidate passes through this same automated screen. The workflow, published as a template on Sim's own platform, separates hiring into two tiers. The first tier is high-volume and repetitive: opening attachments, extracting the same structured facts from each resume, checking each against a requirements list, drafting the same rejection with only the name changed. The hundredth resume does not receive the attention the first one did; that is a property of human attention, not a failure of diligence. Automation applies to this tier. The second tier — the deciding — does not improve by being done faster.

The agent parses every resume for eight standard facts: role scope, tech stack depth, project ownership, deployment scale, open-source contribution, team size managed, shipping cadence, and evidence of agent-orchestration experience. Each fact maps to a weighted criterion the hiring team sets per role. A backend-engineering role weights deployment scale and stack depth; a developer-relations role weights open-source contribution and shipping cadence. The weights are not hidden; they are encoded in the workflow definition that candidates can inspect if they know where to look.

Thresholds control the outcome. The template defaults to a wide band between the automatic-rejection floor and the automatic-advance ceiling. Candidates who fall inside the band are not rejected; they are queued for human review with a drafted summary waiting. A narrow band pushes more candidates into automatic drafting, which shifts the arrangement toward automated decisions. Sim's team keeps the band wide. The review remains daily. Drafts wait for review rather than waiting to be written.

This design reflects a constraint Sim's founders observed across 100,000 builders using the platform: reading compresses; deciding does not. The agent processes the full inbound volume, roughly 158 runs of the screening workflow at a 95 percent success rate, Simular AI's data shows, on the template's benchmark, but the final invitation to interview stays human. The goal, as the workflow documentation puts it, is to screen more people without being constrained by reviewer capacity, while giving reviewers more time to build relationships with candidates.

For applicants, the implication is concrete. The screen rewards structured evidence over narrative flair. A bullet that reads "designed and deployed a multi-agent RAG pipeline serving 12k daily active users on Kubernetes" extracts cleanly. A paragraph about "passion for building scalable AI systems" does not. The agent looks for the eight facts; if they are absent, the candidate lands in the rejection tier regardless of pedigree. If they are present but thin, the candidate lands in the review band, where a human reads the drafted summary and decides whether to advance.

The workflow also flags applications that appear to have been generated by a large language model. The template includes a detection step that highlights sections with low perplexity and high token-probability uniformity, patterns common in LLM-written cover letters and resumes. Flagged applications are not auto-rejected; they are routed to the review band with a note. The signal is advisory. Reviewers weigh it alongside the extracted facts.

Candidates who understand this architecture optimize for extractability. They lead with the eight facts in the first third of the resume. They quantify deployment scale, name the orchestration framework, link the GitHub repo. They avoid prose the parser cannot map. They do not try to game the detection step; they write like a human who has shipped agentic workflows, because the next stage is a conversation with the engineers who built the parser.

Seven Roles, One Screen

Sim, the San Francisco–based Series A startup building an open-source platform for agentic AI workflows (often described as "Figma for AI agents"), lists five open positions on its public job boards as of August 2026. The company's own careers page and third-party aggregators (TheJobsMap, FastAIJobs) consistently show the same five roles; two additional requisitions are not publicly posted. What follows covers the five documented roles.

Sim's platform combines a visual workflow editor with knowledge bases, structured data tables, file storage, and execution logging, enabling developers to go from a single agent to complex multi-agent systems in production. The company reports 100,000+ builders, 1.27 million hours of work completed, and $20 million in user savings, backed by a $7M Series A closed in November 2025. With a total headcount of roughly six, each hire represents a material percentage of the team.

Founding Account Executive: The first dedicated sales hire. This role owns the full cycle from inbound qualification to close, targeting technical buyers at companies adopting AI agent workflows. The compensation band ($150k–$180k base, presumably with variable, according to Sim's careers page) reflects an early-stage GTM motion where the founder still carries quota. Candidates need experience selling developer tools or infrastructure to engineering leaders, not just procurement.

Founding GTM Engineer: A hybrid technical/GTM role sitting between product and sales. The band ($210k–$260k, Sim's careers page reports) signals seniority: this person builds demos, writes integration code, and partners with prospects to prove value in their environment before a contract lands. The "founding" prefix means defining the playbook, not executing an existing one. Expect to ship code that becomes part of the pre-sales motion.

Software Engineer, Product: Core product engineering on the workflow canvas, execution engine, and developer experience. The stack centers on TypeScript/React for the frontend, with a Rust or Go backend for the agent runtime. This role ships user-facing features — think drag-and-drop node composition, real-time collaboration, version control for agent graphs — and owns them end to end.

Software Engineer, Agent Harness: Deeper in the runtime: the orchestration layer that schedules, monitors, and retries agent steps; manages state across long-running workflows; and exposes observability hooks. This is systems work — concurrency, durability, sandboxing — not UI. The same $210k–$260k band applies, reflecting the specialized demand for engineers who understand both LLM tool-calling patterns and distributed systems primitives.

Design Engineer: The broadest band ($120k–$200k) suggests flexibility on seniority. This role bridges design and frontend implementation: building the component library, refining the canvas UX, and ensuring the editor feels responsive at 10,000-node graphs. Proficiency in React, CSS-in-JS or Tailwind, and a portfolio of complex interactive UIs are table stakes.

All five roles are San Francisco–based (no remote tags appear in the listings). The concentration of three roles in the $210k–$260k band — two in engineering, one in GTM — shows where Sim is investing: technical depth in the agent runtime and a sales motion that requires engineering credibility. The two "founding" titles on the GTM side indicate the company is still defining its market approach; the person who takes either seat will write the playbook the next ten hires follow.

What a Portfolio Must Show

Sim's public footprint (its open-source repository, its product metrics, and the workflows its founder demonstrates) signals what the screen likely rewards. The company's GitHub repository (simstudioai/sim) shows 7,232 commits, 29,435 stars, and 3,778 forks as of September 2026. That activity level puts it in the top tier of agentic-AI frameworks.

A September 2026 YouTube walkthrough by builder Kyle Behrend, an AI, Automations and Systems Specialist, describes the output: "by the time I sat down to actually look at this deal, there was already a full research report on the company, a safety audit telling me they were fine to work with, and three different video concepts that I could choose from in a small presentation that I could review in under 2 minutes." Behrend calls out "mission control and the orchestration" as the standout features, phrasing that maps directly to the capabilities industry analysts identify as core to agentic infrastructure: dynamic discovery, agent-to-agent collaboration, persistent memory, and event-store auditability.

What follows is inference from product to screening logic. Sim builds the runtime and orchestration layer for agent teams. Its screen therefore filters for evidence that a candidate has shipped — or at least architected — systems where multiple agents share context, call tools, and persist state across steps. A single-agent chatbot demo will not pass. The repository history shows heavy TypeScript and Python usage; a portfolio that includes typed agent definitions, tool schemas, and eval harnesses for multi-step workflows aligns with the codebase.

GitHub evidence carries weight. A fork of simstudioai/sim with meaningful contributions (bug fixes in the orchestration loop, new tool integrations, memory-backend swaps) signals deeper engagement than a starred repo. Public repositories that implement Model Context Protocol servers, agent registries, or evaluation pipelines for agent teams read as direct proof of the competencies Sim hires for. Candidates who have built against those standards demonstrate the portability mindset Sim's architecture demands.

Design candidates face a different bar. The product surfaces a visual mission-control dashboard where operators monitor agent fleets, inspect decision traces, and intervene. Portfolio pieces that show complex state visualization (branching execution trees, token-usage heatmaps, cost-attribution timelines) speak the same language. Go-to-market applicants should demonstrate they can translate agent capabilities into buyer outcomes: Behrend's "saved me a bunch of time" is the unit of value Sim sells.

No public document from Sim enumerates a portfolio checklist. What exists is a consistent pattern: the product rewards builders who treat agents as reusable, auditable, composable services. The screen almost certainly mirrors that pattern. Candidates who submit a single polished demo of a three-agent workflow with persistent memory, tool use, and an eval report will have more signal than ten disconnected prompt-engineering notebooks.

Inside the Interview

Candidates who clear Sim's automated screen enter a multi-stage interview process designed to test the same AI-agent competencies the screen filters for, only now with human evaluators probing depth, judgment, and communication. The company has not published a public interview playbook, but its technical focus (agentic engineering workflows, browser-native simulation at scale) and the documented patterns at peer firms hiring for similar roles make the contour of the process predictable.

Technical assessment: code, architecture, and AI-specific judgment

Expect a take-home or live coding exercise centered on the problems Sim actually solves: orchestrating simulation setup, design-space exploration, and reporting through agentic workflows. Anthropic's performance-engineering team, which hires for comparable GPU-native, AI-accelerated workloads, has run a take-home where candidates optimize code for a simulated accelerator (over 1,000 completions since early 2024, with dozens of hires including engineers who brought up Trainium clusters and shipped every model since Claude 3 Opus). Their v3 puzzle uses a tiny, heavily constrained instruction set to force out-of-distribution reasoning that current models struggle to replicate under time pressure. Sim's equivalent will likely test whether you can reason about simulation pipelines — mesh generation, solver configuration, result post-processing — not just call LLM APIs.

System design carries the heaviest weight in AI engineering interviews across the industry. Interviewers will ask you to architect a RAG pipeline for retrieval-augmented simulation guidance, or an agent that plans and executes multi-step CFD studies. The research is consistent: RAG is the most common architecture pattern in 2026, and agentic AI is the fastest-growing area. You should be ready to defend choices on chunking strategy, embedding models, reranking, and guardrails against hallucinated boundary conditions; hallucination management is a core production concern, and interviewers want to see it treated as an engineering problem, not an unsolvable mystery.

Cost-at-scale questions appear routinely. AI features can become extremely expensive at scale; this tests whether you think about the business side of AI engineering. Sim's per-minute GPU pricing model makes token efficiency and inference optimization directly relevant to revenue.

Behavioral and culture fit: the human layer

Expect interviewers to probe how you explain complex technical trade-offs to non-technical stakeholders (product managers, sales engineers, customers running their first simulation). Many candidates fail AI engineer interviews not because they lack technical skill, but because they can't articulate why they made specific decisions.

The behavioral round will test whether you can do the part the model can't (synthesize context, read the room, and adapt the narrative).

What to prepare

• Walk through one end-to-end AI feature you shipped: not a demo, a production system. Be specific about evaluation strategy (traditional ML metrics don't translate cleanly to generative AI), latency budgets, failure modes, and how you measured hallucination rates.
• Practice system design out loud: RAG for simulation knowledge bases, agent orchestration for design-space exploration, multi-modal inputs (geometry + boundary conditions + historical results).
• Prepare behavioral stories demonstrating AI-specific judgment: a time you chose not to use an LLM, a time you caught a hallucination before it reached a user, a time you optimized cost without degrading quality.
• Know Sim's domain: CFD, FEA, electromagnetic and thermal simulation, the Lumi AI assistant that handles mesh generation automatically, the shift toward physics AI and agentic workflows that "orchestrate simulation setup, execution, design space exploration, and reporting."

The screen filters for keywords and evidence. The interview filters for the reasoning behind them.

The Team Behind the Platform

The research on Sim's internal team is thin (no founder interviews, no public org chart). What the product and its traction do reveal, however, is a team that ships at a pace and breadth that constrains what the composition must look like.

Founded in 2025, Sim has accumulated 7,232 commits and 29,435 GitHub stars. That commit volume across a codebase spanning a visual workflow canvas, a semantic knowledge base, structured tables, a shared file store, execution logs with cost tracking, 1,000+ pre-built integrations, multi-LLM routing, three authoring paradigms (visual, conversational, code), and enterprise controls (SSO/SCIM, permission groups, spend limits, audit records, data retention, self-hosting via Docker/Kubernetes, SOC 2 Type II, ISO 27001, GDPR) implies a founding engineering team that is both full-stack and infrastructure-fluent. You do not build a self-hostable, Apache-2.0-licensed agent runtime with provider-agnostic model support and per-run cost attribution without engineers who have wrestled with container orchestration, authentication systems, and the messy reality of third-party API reliability.

The named customers (Rivian, Russell Investments, eXp Realty) signal a team that has already navigated enterprise procurement, security reviews, and deployment in regulated or data-sensitive environments. Behrend's testimonial calls out "custom roles with specific permissions, skill selection, custom support and docs" (features that exist because someone on the team anticipated the governance needs of a 5-seat team and a 500-seat organization simultaneously).

That is not a feature you prioritize unless the team itself operates that way (unless the people building Sim include both the operator who thinks in graphs and the engineer who thinks in functions, and they have already resolved the friction between those modes internally). The same logic applies to the knowledge base and tables: semantic memory over proprietary data and structured storage that agents read and write across runs are solutions to problems the team has felt themselves while moving prototypes to production.

The competitive landscape (n8n, Dify, Langflow, Flowise, CopilotKit, SuperAGI, and every model vendor building its own builder) means the team is optimizing for differentiation on control and observability, not just orchestration. The observation that "logs with cost tracking are the feature you will care about in month three" reads like a scar, not a spec. Someone on the team has watched an unattended agent rack up a bill that needed explaining to finance.

What the team prioritizes in hires can be reverse-engineered from what the product demands: engineers who have integrated dozens of SaaS APIs and know which ones lie about rate limits; designers who can make a visual graph editor feel like a tool, not a toy; product-minded builders who understand that "delegating real work to agents" means the platform must handle permissioning, auditability, and cost governance from day one. The open-source core (29,435 stars, 3,778 forks) suggests they value contributors who can operate in public, accept external PRs, and maintain backward compatibility across a moving target of LLM APIs.

The research does not name the founders or early employees. But a team that delivers a SOC 2 Type II certified, ISO 27001 compliant, self-hostable agent platform with 100,000+ builders and $20M in documented customer savings within roughly 18 months of founding is a team that hires for shipping scope, not narrow specialization. If you are applying, your evidence should show you have built across the stack (auth, infra, integrations, UX) and that you have felt the pain of an agent running wild in production.

The gatekeeper agent that read your resume was built by people who have felt that pain. They will be the ones across the table when the screen passes you through.

Working in AI? Zero G Talent tracks the openings: see every open Databricks role, browse AI jobs, openings at Anthropic, and the people building the field.