Tenor's $150K-$200K AI Role Filters Out Résumés, Requires One Artifact

The Screen: One Artifact, No Cover Letter

Tenor has eliminated the cover letter and the résumé. One artifact, one role, two minutes. That is the entire gate.

The careers page states the mechanic plainly: "Send one artifact. Pick one role. Skip the cover letter. Applying takes less than two minutes." No recruiter screen. No take-home assignment. No portfolio review. The artifact is the screen.

Candidates choose among three Member of Technical Staff openings: Systems Engineering (runtime, tools, verification), Context & Memory (context, retrieval, memory), and Evals & Task Engineering (tasks, evaluation, quality), plus a "Build Your Own Role" option. All are San Francisco–first with remote considered for exceptional candidates; all are full-time with equity and bonus.

The hiring signal is explicit: "This is not a normal coding job." The page lists what disqualifies you — "You want narrow tickets and clean boundaries," "You like demos more than reliability," "You avoid understanding the system," "You would rather look productive than prove what changed" — and what passes: "Own difficult systems." The company's LLM context page declares "proof over theater." The careers page demands "Prefer shipped evidence over process theater." The artifact is the only evidence that counts.

Tenor lists one role added in the past seven days on Zero G Talent's board: Founding Engineer, Agent and Data Systems (San Francisco, CA, US), USD 150K–200K/yr. The careers page lists compensation as "Competitive compensation · equity · bonus" without a published band for the Member of Technical Staff roles. The negative filters make the bar clear: a submission that looks like a résumé, a demo, or a scoped ticket will not clear the screen. The artifact must demonstrate ownership of a difficult system end to end, with logs, state, tests, traces, and user behavior as verification.

Which Tenor Is Hiring?

Two companies share the Tenor name on job boards, and the overlap is already confusing candidates.

The Zero G Talent board lists Tenor with a single live opening: the founding engineer role in San Francisco at $150K–$200K. That Tenor builds AI employees — agent runtimes, tool harnesses, memory layers, verification loops — the infrastructure that lets autonomous systems operate reliably in production. Its public homepage lists team backgrounds including Stanford, UC Berkeley, Technical University of Munich, International Mathematical Olympiad, Amazon, Tesla, and Waymo.

A different Tenor (tenorhq.com) announced a $5.4 million seed round on December 19, 2024, led by Base10 Partners with participation from Reach Capital, GlobeNewswire reported. Its co-founders are Charlie Stigler and James Cross, both former founders whose previous startups were acquired by Workday and now underpin Workday Learning. This Tenor sells an AI leadership-development platform: voice simulations for practicing difficult conversations, coaching feedback in Slack and Teams, analytics tied to HR systems. Its careers page at jobs.ashbyhq.com/tenor/ shows a Sales Development Rep role, remote across the U.S. and Canada. The product is built for L&D buyers, not for engineers shipping agent infrastructure.

The confusion is structural. Aggregators scrape both domains into the same "Tenor" bucket. A candidate searching "Tenor AI jobs" sees a founding-engineer role in San Francisco next to a sales role tagged "remote (US/Canada)," with no signal the companies are unrelated. The leadership-development Tenor's funding announcement got pickup in HR-tech outlets; the agent-systems Tenor has not issued a press release. That asymmetry means the leadership-development narrative dominates search results, and the engineering role looks like an outlier — or worse, a ghost listing.

The disambiguation matters because the application mechanics differ entirely. The leadership-development Tenor hires through a conventional funnel. The agent-systems Tenor (the one on the Zero G Talent board) requires a single artifact: a link to a repository, a deployed system, or a technical write-up proving you can build the harness, runtime, tooling, memory, and verification layers the role demands. No cover letter. No résumé. Applications are sent by email to a founder address.

If you are an engineer targeting the Founding Engineer role, verify you are on the right company page. The Zero G Talent listing links to Tenor with the San Francisco location. The leadership-development Tenor's founders are Stigler and Cross; its product is a coaching platform; its open role is sales. They are not the same company, they do not share a codebase, and they do not share a hiring process. Submitting a résumé to the agent-systems Tenor is not just ineffective — it is the exact signal the screen filters out.

Salary, Geography, and the Seniority Signal

Tenor's board listing for the founding engineer role posts a base salary band in San Francisco. The title — "Founding Engineer" — carries the weight. At a company with one salaried role on the board, this hire is effectively the first technical employee. The narrow band width reinforces that: no ladder yet, no leveling matrix, no room to negotiate a "Staff" vs "Senior" grade. You join at the founding tier or you don't join.

Geography sharpens the read. The role is pinned to San Francisco, CA, US — not "Bay Area," not "hybrid flexible," not "remote optional." That specificity signals an onsite default. For an agent-systems role touching those layers, physical proximity to the founders and to GPU clusters or eval infrastructure matters. The location constraint filters for candidates who have already made the Bay Area bet, or are willing to make it, which correlates with higher risk tolerance and tighter networks to the frontier-AI ecosystem.

Those roles are effectively contracted model-training labor at scale, not founding-team systems ownership. Tenor's band is lower because the asset class is different: you are not selling labeled hours; you are buying into a cap table. A similarly titled SVP search at a private-equity portfolio company drew 200-plus applicants, underscoring that cash-rich, later-stage transformation roles attract volume. Tenor's applicant count is not public; the artifact gate is designed to keep that number low and the signal high.

The seniority signal, then, is not in the base number. It is in the combination: a founding-engineer title, a single-digit equity grant, a San Francisco anchor, and a screen that demands shipped agent-systems artifacts before a human reads a line of your background. That package selects for builders who have already internalized the proof-of-work standard and who can afford to bet their next year on it.

What the Artifact Actually Has to Prove

Tenor's artifact screen does not want a prompt library. The role sits at the harness layer, where research converges on a single thesis: autonomous software-engineering capability emerges from the model–harness–environment system, not the model alone. A winning submission proves you can build the runtime substrate that turns latent model capability into auditable engineering behavior.

The harness carries eleven responsibilities: task specification, context selection, tool access, project memory, task state, observability, failure attribution, verification, permissions, entropy auditing, and intervention recording. Your artifact should demonstrate at least three working together. A repo showing context selection feeding a tool call, which updates project memory, which then informs the next context selection: that is a harness loop. A notebook that calls an API and prints the result is not.

Runtime matters more than the model name. NVIDIA's NOOA results make this concrete: the same GPT-5.5 model scored 82.2% on SWE-bench Verified with 29 LLM calls and 1.1 million tokens, while comparison harnesses needed 66 calls and 2.2 million tokens to reach 78.2%. Harness design alone accounted for double-digit swings. Your artifact should show you understand why: pass-by-reference tool results, programmable loop engineering, explicit object state, model-callable harness APIs. If your submission treats the model as a black box you prompt, you have not built a harness.

Memory is not a vector store. The Augment Cosmos guide notes that an LLM carries no memory between invocations, so systems must externalize state. A winning artifact demonstrates cross-session persistence: task state preserved across days, project context that survives process death, skill files that version independently of session context. The UNU framework distinguishes project memory from temporary session context; mixing them is a documented failure mode. Show the separation.

Verification replaces trust. "Don't ask the agent if it succeeded; ask it to leave a receipt" — the CoreWeave team's gate treatment recorded 20 of 24 safe-and-useful outcomes with receipts for what it blocked and allowed. Your artifact needs a verifier that runs independently of the agent loop, produces inspectable evidence, and gates continuation. A test suite the agent itself writes and runs does not count. The receipt must come from a separate, cheaper decision model or deterministic check.

Tools need boundaries. The UNU report lists tool blast radius and credential mishandling as harness-layer risks when agents access shells, files, message channels, or APIs without sufficient scoping. A submission that hands the model a raw shell wrapper fails. Show least-privilege tool governance: typed interfaces, approval gates for state-changing actions, sandbox isolation, and observability that distinguishes a healthy server from an agent quietly hallucinating actions.

Observability is not logging. Traditional APM fails because agents are non-deterministic and can fail silently; latency looks fine while the agent hallucinates. Your artifact should surface failure attribution (which component caused the deviation), intervention recording (where a human stepped in), and entropy auditing (where the run drifted). NVIDIA's NOOA architecture includes an entropy dashboard; the arXiv framework calls for entropy dashboards as AI-native repository affordances. Build one.

The artifact itself is the interview. Tenor is hiring at a founding-engineer band at that level in San Francisco. That compensation signals they need someone who can ship the harness layer end-to-end, not iterate on prompts. The screen filters for the engineer who has already felt the pain of a process dying mid-loop and losing the run, who has built the recovery procedure, who has watched a plausible output pass the model's self-check and fail in production. Ship the system that prevents that. The résumé never enters the room.

Conflicting Trackers and Ghost-Job Risk

Tenor's board footprint is unambiguous: one role, posted within the last seven days, carrying a $150K–$200K band and a title that names the stack and the mandate. That specificity alone separates it from the "rockstar wanted" templates that dominate evergreen funnels. But step outside Zero G Talent's first-party feed and the picture fractures.

PulseIndex.ai counts open roles across 56 public companies, refreshed daily from career pages. Tenor, a private startup, does not appear in that universe. The McCoy Hiring Tracker rolls up Fortune 100 ATS feeds including Alphabet and Berkshire subsidiaries; it has no reason to surface an early-stage agent-systems shop. The Big Tech Hiring Tracker monitors 22 name-brand employers and 1,565 live roles; Tenor is not on that list either. None of these trackers are built to catch early-stage AI companies hiring their first founding engineer. Their silence is a coverage gap, not a signal the role is fake.

The ghost-job heuristics tell a different story when applied directly to the listing. Posting age: under seven days, well inside the 14-day median close window RemoteOnly.io documented for remote roles and far below Alion's 62-day median live-posting age (Sept 2026). Salary transparency: a declared $150K–$200K band satisfies pay-transparency laws in California, New York, Colorado, and Washington, jurisdictions where missing ranges cut conversion rates in half. Specificity: the title names "Agent and Data Systems," the location is San Francisco, and the artifact-first screen is described in the posting. Re-post count: no evidence of recycling; the board shows a single addition in the trailing week.

A listing that is a month old on a board that never re-checks is more likely closed than open.

That line from RemoteOnly.io's detector captures the core risk: stale boards, not stale roles. Zero G Talent ingests directly from the source ATS, so the "added in past 7 days" timestamp reflects the employer's actual publish event. Third-party aggregators often lag by days or weeks, and some scrape once then stop checking. A candidate who only checks LinkedIn or Indeed may see a phantom copy of the role, or miss it entirely, while the canonical version sits live on the company's own careers page and on the board that ingested it natively.

Broader market data reinforces why this distinction matters. Greenhouse's 2025 survey found 18–22% of 2024 postings never filled; three in five candidates suspected a ghost job. Clarify Capital put the >30-day active rate at one in seven. Alion's index flagged 16.5% of postings with ghost signals and 39.9% open >90 days. LastRound AI measured 21.8% of Lever postings open >6 months. But those are aggregate rates across millions of listings, many from the 1,001–5,000 employee cohort that generates 24.8% ghost rates, per the Interview Guys' compilation. Tenor sits outside that cohort entirely.

The conflict isn't between trackers. It's between trackers that don't cover this segment and a first-party feed that does. The ghost-job risk for this specific role is low: fresh, specific, salaried, and transparent. The real risk is a candidate trusting a tracker that hasn't indexed Tenor at all and assuming the role doesn't exist.

The Broader Shift to Proof-of-Work AI Hiring

The application flood is real and measured. U.K. graduate employers fielded an average of 140 applications per vacancy in 2025, the highest level in three decades, the Institute of Student Employers reports. In the U.S., the Financial Times found about half of job applicants now use AI to write cover letters and keyword-optimize résumés, roughly doubling applications per posting. The result is a signal-to-noise collapse: Belkins analyzed 7.5 million cold emails across 2025 and found a 0.45% reply rate. Founders and owners replied at 0.57%; companies under 10 employees hit 0.72%, more than three times the enterprise rate. Hiring managers are drowning. A 2026 Resume Genius survey of 625 U.S. managers found 36% spend under 30 seconds on a cover letter, 48% spend 30 seconds to two minutes, and only 15% go beyond two minutes. Yet 94% said cover letters influence interview decisions. The document survives, but the attention to read it does not.

Companies are responding in three ways. First, skills-based hiring: roughly two-thirds of U.K. organizations have adopted it, Michael Page's 2026 research shows, and 60% of candidates say they prefer roles defined by skills over credentials. Second, practical assessment. Taktile screens for interpersonal skill, self-reflection, and a willingness to question what a candidate has built. PMG puts graduate candidates through in-person exercises, team work, and live client briefs. Third, explicit anti-AI gates. Anthropic attached an "AI Policy for Application" to all 150-plus open roles: "Please do not use AI assistants during the application process. We want to understand your personal interest in Anthropic without mediation through an AI system, and we also want to evaluate your non-AI-assisted communication skills." Open-source developer Simon Willison spotted the policy. Dumb.co went analog: after listings went live, they required handwritten cover letters answering specific prompts about "dumb living." Nearly 900 people applied for four internships; 70% submitted handwritten letters, and 200 to 300 physically mailed them. Director of brand and communications Afreka Ebanks said the letters "weren't connecting with our mission and our greater goal" when typed; the handwritten ones revealed attitude and passion. Intern Ian-Karlo Krockel, 21, said he had no luck elsewhere until Dumb.co: "It felt like someone actually was looking through what I had to say and that it wasn't just going through some sort of filtering software."

Tenor's artifact screen sits at the sharp end of this spectrum. The role asks for one artifact, no cover letter, no résumé. The compensation band signals a senior builder who can ship agent systems on harness, runtime, tools, memory, and verification. The founders are betting that a demonstrated system beats a credentialed résumé every time. That bet aligns with what Taktile's Broadhead told CNBC: "The pace of model releases is so fast that the ceiling of what you can do changes every three months. And so the question we're asking is: what have you changed about your AI tools in the last three weeks?" Krichefski put it more bluntly: "You don't really know whether someone's going to have the skills to do something until you challenge them to do it. The past doesn't tell you that."

The shift is structural. Judd Kessler described the current loop as "AI selecting AI to elevate to humans for review." He predicts both sides will reach for old-school proof (letters of recommendation, phone calls, trusted referrals) because they depend on human relationships AI cannot fabricate. Tenor's artifact is a different old-school proof: code that runs, systems that hold. The cover letter, as angld.ai argues, is built for a recruiter stack, not a busy person on a phone. Its replacement isn't a shorter letter; it's a different artifact with a different goal — a reply, not a hiring decision. Tenor makes that replacement explicit. The artifact is the decision.

How to Clear Tenor's Gate

Tenor's application page repeats the rule described earlier: submit one artifact for one role, no cover letter. The screen filters out candidates who lead with credentials instead of proof. Here is what a submission needs to survive it.

Pick exactly one role and tailor the artifact to it. ** As noted earlier, the careers page lists those three roles: Systems Engineering Runtime, Context & Memory, and Evals & Task Engineering, plus that same option. Each demands a different slice of agent-systems work. A runtime artifact shows you can build harnesses, tooling, and verification loops. A context-and-memory artifact demonstrates retrieval, state management, and long-horizon reasoning. An evals artifact proves you can design tasks, measure quality, and close the loop between signal and improvement. Submitting a generic "AI portfolio" across multiple roles signals you haven't read the brief.

The artifact must be a shipped system, not a demo. As the page requires, favor shipped evidence over process theater. "Move fast, then verify with that same evidence." A winning submission shows a system that ran in production or a realistic staging environment long enough to generate observable behavior — logs, traces, eval results, user feedback — and it shows what you changed when the data contradicted your assumptions. A notebook that runs once on a toy dataset does not meet this bar. A GitHub repo with a clean README, reproducible runs, and a changelog documenting decisions under uncertainty does.

Own the system end to end. The role descriptions emphasize runtime, tools, verification, memory, retrieval, and task engineering as integrated concerns, not isolated tickets. Your artifact should reflect that scope: you defined the interface, chose the infrastructure, instrumented the critical paths, and iterated when the evals regressed. If you only touched the model layer while someone else handled deployment and monitoring, say so, and explain what you would have done differently had you owned the full stack.

Show judgment, not just tooling. "Use the best tools without outsourcing judgment" appears on the careers page. The artifact should make visible where you made a call a library or framework couldn't make for you: choosing a retrieval strategy, setting an eval threshold, designing a memory consolidation policy, deciding when to escalate to a human. Document the alternatives you considered and why you rejected them.

Prove what changed. The careers page demands that you take ownership of difficult systems, make clear decisions, and prove what changed. The artifact needs a before-and-after narrative grounded in metrics: latency dropped from X to Y after you rewrote the tool-calling loop; eval pass rate rose from 62% to 84% after you redesigned the task decomposition; memory hallucinations fell by half after you introduced a verification pass. Screenshots of dashboards, links to wandb or MLflow runs, or a short write-up with numbers all work. Vague claims of "improved performance" get filtered.

Skip the cover letter entirely. The instruction is explicit. Any energy spent crafting a narrative about your passion for AI is energy not spent sharpening the artifact. The screen reads the artifact; it does not read your motivation statement.

San Francisco first; remote for exceptional candidates. The location constraint is real. If you are not in the Bay Area, your artifact must be unambiguously exceptional: the kind that makes a hiring manager argue for an exception. Plan accordingly.

The Gate Keeps Itself

The artifact screen does not scale. It cannot. A founding engineer who has built a harness loop that survives entropy, a memory layer that persists across sessions, a verifier that catches what the model misses — that engineer proves the work by showing the work. Tenor's founders built the screen to find engineers who already operate at that scope. The artifact is your proof that you do.

Working in frontier tech? Zero G Talent tracks the openings: see every open Tenor role, browse frontier tech jobs, the companies hiring, and the people building the field.