The Roles and the Screen Behind Them
Hercules’ public Ashby board lists 24 openings, with staff-level engineering and product work at the center of a company-wide hiring push. The public briefs make the hiring criteria unusually clear: from agent evaluations and production infrastructure to customer-facing workflows, they reward evidence of work that reached customers rather than pedigree alone.
Hercules says on its careers page that it is a seven-person AI app builder for business, backed by a16z, with tens of millions in revenue. It says hundreds of thousands of businesses have used Hercules to build custom CRMs, ERPs, HRIS tools, client portals, inventory management systems, and more.
Hercules’ founders were part of Fig.’s founding team. TechCrunch reported in 2023 that Amazon acquired Fig’s technology and that Fig employees, including two co-founders, joined AWS. Hercules says Fig was acquired by AWS and became Amazon’s AI coding assistant, Kiro CLI. The openings use the language of agents, evaluations, and model access, while other roles cover developer workflow and production infrastructure. The cited AI-agent posting does not list prior service at Fig or AWS as a qualification.
Zero G Talent’s current snapshot names 20 of the 24 openings, which can be grouped as follows.
The staff titles show the level of ownership Hercules seeks: all eight engineering openings carry staff titles, and both product-management titles are staff-level.
What the screen specifies
Hercules’ public materials do not name an interview sequence, take-home exercise, or timed test. Anyone who claims to know the exact loop or hidden scorecard would be guessing. The job briefs reveal a defensible structure, and the AI-agent role makes it clearest.
Ownership comes first. The listing states: “Build, maintain, and own Hercules' AI coding agent”. That mandate carries operating responsibility rather than a request for research commentary.
Evaluation belongs in the work. The role requires the hire to make the agent cheaper, faster, and more accurate. A candidate who can describe prompts but not failure analysis, testing, or trade-offs has addressed only part of the brief.
The work runs at production scale. Hercules said the hire can draw on more than 1 million repositories and millions of chat messages and can use models from OpenAI, Anthropic, and Google. The scale tests the ability to operate an agent around real usage, not a toy exercise.
Shipping is the stated behavior. The listing puts it plainly:
“We ship. We ship fast, iterate, and correct mistakes quickly.”
The same posting said that industry best practices change every week and that Hercules still wants someone who has “wrestled with these challenges at scale and has strong opinions about what works and what doesn't”. The wording asks for hard evidence rather than a list of fashionable tools.
Together, the brief asks candidates to own a system, measure it, learn from failure, and keep the work moving. That is a test of operating judgment.
The Hiring Burst and What It Signals
Hercules lists its salaried roles on Zero G Talent, and the company added one opening in the past seven days. The latest is the Staff Technical Product Manager, Evals, a product role focused on model evaluation.
The Zero G Talent board records when roles are added to its tracker; that does not establish when every posting first appeared. The one-new-role figure does not show that all 24 positions arrived within a week, and calling the slate a simultaneous hiring sprint would exceed the record.
The scale is material. Hercules’ own careers page puts the company at seven people. Against that base, 24 live roles describe a broad expansion, but an opening count is not a published headcount target: the materials do not say every position will be filled as a net addition.
More than half of the 20 roles named in the table sit in engineering/data and product: 11 roles. The slate is not confined to builders: it also covers the commercial and support work required to put a platform in front of customers.
Name and funding records need care. A separate legal-AI company, HerculesAI, announced a $26 million Series B. That round is not funding for the app builder at hercules.app and should not be used as evidence for this hiring plan. The app builder’s own site says it is backed by a16z.
Hercules says businesses use the platform for the customer applications described above. Its job mix mirrors that scope: the agent, evals, billing, data, infrastructure, growth, support, and sales.
The company’s product and commercial ambitions are landing in the labor market at the same time. The shape of the slate makes geography the first filter for many applicants: most named roles are anchored in San Francisco.
San Francisco First, Remote Second
In the October 2026 Hercules snapshot, the location labels divide sharply between San Francisco and remote work.
| Location label | Named openings |
|---|---|
| San Francisco | 17 |
| Remote | 3 |
| Total with a location label | 20 |
Because the table covers only records with usable titles and location labels, it is not a final on-site/remote classification for every opening. For an applicant outside the Bay Area, geography is an immediate filter alongside technical evidence.
The role map above shows the pattern: engineering/data, product, content, and the two named account-executive openings carry San Francisco tags, while the EMEA and APAC support roles and the growth role are remote. The exceptions cluster in support and growth; no named staff software-engineering or staff technical product-manager role is remote. Hercules’ own careers site lists the AI-agent opening in the same city, corroborating the board.
For the AI-agent role, Hercules says its founding team works in-office in San Francisco, around Kearny and Bush, and strongly prefers in-office work. It says remote candidates must be exceptional, work in US time zones, and travel to the office regularly.
The policy and role-level listings do not align perfectly. The general remote rule points to US time zones, whereas the two regional support openings and the growth role extend well beyond the United States. Candidates should rely on the geography printed on the specific role rather than the general preference.
Hercules is also recruiting into a crowded Bay Area market. An AI Engineer Hiring in San Francisco Bay Area market report, updated in June 2026, shows how much tighter demand has become:
| Bay Area measure | Result |
|---|---|
| AI-engineer job demand | Nearly one-quarter more than a year earlier |
| Senior-level openings in Q1 2026 | 4,800 |
| Q1 2026 total vs. all of 2020 | Twice as many |
Those regional figures do not count Hercules openings, but they show the pressure Hercules faces as it recruits in a concentrated Bay Area labor market. That pressure makes an office-first model more consequential, not less.
Before submitting, answer four location questions: Where can you base? Which time zones does the role accept? Does the role allow remote work? How often can you travel? That geography check should be made alongside the machine-readable résumé and any recorded assessment.
The Candidate Reaction: Decoding a Black-Box Screen
At the application stage, software screening enters the picture. A JobHire analysis updated in July 2026 found that 97.8% of Fortune 500 companies use an applicant tracking system on their career page and 99.7% of recruiters use keyword filters within those systems.
Hercules does not identify its screening vendor, so calling its system a HireVue setup would exceed the evidence. It also does not publish details of a recorded interview. The wider research shows what applicants may encounter elsewhere: a machine-readable application, often followed by a machine-scored interview.
In the systems JobHire describes, the process begins with the résumé. A submitted file does not simply land in a recruiter’s inbox. An applicant tracking system extracts the text, structures the information, and ranks the candidate. Multi-column layouts, tables, text boxes, and decorative graphics can garble the record. Contact details in headers or footers may disappear. Without a dedicated skills section, the parser has less organized material to match against the job description.
A polished résumé can therefore be functionally invisible. The JobHire guide recommends a single-column Word document or plain-text PDF, standard fonts, and no tables, decorative images, or hidden contact details. Its quickest quality-control test is blunt: select all the résumé text, paste it into a plain-text editor, and check the reading order. If the email address vanishes or the experience jumbles, the parser sees the same mess. Design the file for extraction, not merely for human readers.
Keyword matching adds another layer. The guide recommends a keyword-match rate of at least 75% with the job description without copying the description wholesale. It warns that a perfect copy can trigger spam filters because the system may infer that the candidate pasted the posting. The tactic addresses a literal vocabulary gap: “teamwork” may not match “cross-functional collaboration,” even when the experience is equivalent. Put job titles, named tools, and core skills where the system can find them.
Passing that stage does not guarantee a conversation. In Final Round AI’s account of HireVue, the company said recorded interviews give candidates less than a minute to prepare and a few minutes to answer. Modern Hire can add a cognitive pre-screen, while Spark Hire relies less on automated scoring but still applies an algorithmic threshold for review. HireVue does not accept or reject candidates on its own; it produces a ranked shortlist. Recruiters review recordings above the threshold, while people below it usually receive no human review.
In the HireVue setup described by Final Round AI, the algorithm is an access gate rather than the final judge. That changes what a candidate must accomplish. The task is not merely to persuade a recruiter; the recording must clear a ranker that may never forward it. Recruiters set the threshold in software, so candidates must make their evidence legible to it.
That makes the hiring application a product: the parser sees only what the file exposes, and the recruiter sees only what the threshold lets through.
Some anxiety about these systems targets a feature that no longer exists. HireVue discontinued its facial-analysis tools in 2021 after an external audit by the AI Now Institute found that the facial scoring introduced bias and lacked predictive validity for job performance. The current platforms focus on verbal content, structural signals, and semantic relevance. They do not algorithmically score eye contact or expressions, although a human reviewer sees the recording after a candidate clears the threshold. Advice built around fear of facial scoring points to a retired feature. The live risks lie in parsing, semantic mismatch, and weak answer structure.
The hiring interface can lose qualified candidates before a recruiter sees them. A Harvard Business School study on hidden workers, cited by JobHire, found that 88% of employers say they lose qualified candidates to automated screening. The candidate may have relevant work, but the system cannot identify it, so the recruiter never sees enough to challenge the ranking.
Candidates can counter by converting work into evidence the machine can score. Final Round AI says current interview systems reward answers that address the question directly, begin and end clearly, and cite specific outcomes. It contrasts “I worked on improving the process” with an account of reducing ticket resolution from four days to 18 hours through a triage-scoring system. JobHire recommends the Challenge, Action, Result method, starting with the result rather than scene-setting and keeping the answer brief.
For an AI developer-tools candidate, compress a production feature into that same problem-action-result structure and place its terms where the parser can find them. Concrete work gives the screen something to match; a prestige signal does not.
Preparation also depends on the question format. Some Goldman Sachs divisions use a strengths-based format, where rigidly scripted Situation, Task, Action, Result answers can score worse than an authentic explanation of motivation. Final Round AI describes competency-based questions as more common at technology companies and banks. Candidates should record three answers, listen for unfinished sentences and filler words, and use the platform’s practice mode to check the camera, microphone, and lighting. Most platforms allow a retake, but the advice is to save it for a factual error or an unfinished answer rather than a minor stumble.
The JobHire analysis found that candidates who choose manual review enter a slower queue, often after recruiters fill the role. That makes the automated screen worth decoding rather than ignoring: make the file parseable, mirror the role’s key terms, and practice a few answers. Once an application clears the machine, it must still persuade a human.
What Clears the Screen: Skills, Shipping, and AI Fluency
Clearing an automated filter, when one is used, does not settle fit. Hercules’ own ads supply the next comparison, and candidates can make the evidence easy to inspect.
The strongest proof is a production agent
A production record is more persuasive than a polished demo. The AI-agent role is a useful test case because the agent is part of the product used to build customer applications; a prompt can hide problems that emerge only under real usage.
The context-engineering requirements show how deep the work runs. Hercules asks for experience with summarization and compression, prompt caching, subagents, memory, retrieval-augmented generation, and deferred tools. These choices determine whether an agent remains useful under real traffic.
A candidate can make that judgment visible with a specific case: explain where context was lost, what the team compressed, cached, or delegated, how it measured the change, and what became harder to inspect afterward. The case should end with the operating result, not a list of tools.
Shipping beats a list of tools
The first-party AI-agent posting demands a track record of independently shipping features to customers. The public candidate-preparation guide also asks for comfort talking with users, proposing ideas, and shipping autonomously. A strong application should show the same behavior. A list of frameworks shows exposure; a deployed example gives the hiring team something to inspect.
A public candidate-preparation guide for the AI-agent role makes the practical response explicit: lead with a demo, recording, or deployed example, then explain the experiments, metrics, trade-offs, model providers and libraries, evaluation pipeline, deployment approach, and own development workflow.
A label such as built an agent is not proof. An agent that failed, whose evaluation caught the weakness, whose fix reached users, and whose result was measured demonstrates capability. The broader AI Engineering Field Guide reinforces that point by treating a portfolio as evidence of skills rather than a container for résumé bullets.
AI fluency means judgment, not keyword coverage
One application question asks: “How do you use AI to accelerate your development?” Together, the posting’s unlimited AI-token budget and that question test judgment rather than keyword coverage: can the candidate turn a noisy, fast-changing capability into a repeatable engineering practice?
The strongest answer should show each step. Say which part of the work AI handled, where it failed, what review caught the failure, and why the workflow was worth keeping. The preparation guide’s examples of plan mode, background agents, and coding assistants can prompt the answer, but the candidate must explain the decision behind the tool. A candidate who cannot say what changed in the development process will struggle to improve a customer’s process.
AI fluency must also survive communication. The same guide’s full-stack version calls for communication that is clear, concise, audience-specific, and actionable. Vague descriptions make agent behavior hard to evaluate when engineers discuss architecture, product trade-offs, and failure modes. A candidate who can describe the same system differently to a user, a product partner, and an infrastructure engineer demonstrates a skill the job requires.
Product judgment and software fundamentals close the gap
Agent experience does not replace ordinary software engineering. Hercules asks candidates to write clean, maintainable code, work effectively in monorepos, design systems and architectures, model data, and design APIs. The public candidate-preparation guide’s full-stack brief also names internationalization, authentication, permissions, admin controls, and branding among the features candidates may need to support reliably. A demonstration can hide a weak data model or brittle access controls. A credible case study should not.
The evals role sits alongside the engineering hires in the table above. That pairing points to a need for people who can connect agent quality, evaluation, product decisions, and production code rather than work inside one isolated layer.
A lead with a deployed system gives the parser and the human reviewer the same evidence. The file becomes a record of decisions rather than a catalogue of capabilities. That is the form the brief itself asks for: ownership, measurement, and a result. The broader market explains why that bar is rising even as overall hiring cools.
What This Says About AI Tooling Hiring Right Now
SignalFire’s 2026 State of Talent Report shows how engineering gained share even as overall large-tech hiring fell.
| SignalFire measure | Result against 2019 |
|---|---|
| Engineers’ share of new hires at 12 Tech Majors | 46% in 2019; 55% in 2025 |
| Large-tech engineering hiring | 11% lower |
| Large-tech total hiring | 25% lower |
| Early-stage startup engineering hiring | 7% higher |
Large-tech engineering hiring gained share while its absolute level contracted. Early-stage startups moved in the opposite direction, hiring more engineers over the same period. That makes hiring a selective expansion inside a broader pullback, not a universal tech boom. Challenger, Gray & Christmas said the tech sector posted its highest single-month layoff total in years in May 2026, with AI the most-cited reason.
Employers are hiring for specific capabilities while cutting elsewhere. That is the market Hercules is hiring in. Its salaried roles target specific capabilities, not proof of a broad recovery.
Current AI-company hiring tells the same story. Migrate Mate’s October 21, 2026 snapshot tracks 3,116 live openings across 10 AI companies, with software engineering the largest listed function at 75 roles. The AI Engineer job board shows where that demand lands. GitHub’s Code|AI team applies research and data to Copilot and VS Code. Greptile uses AI to review pull requests for bugs and anti-patterns. Those tools address the code itself, while Meticulous works on the safety checks that let teams ship AI-generated code and Keycard builds the identity and access infrastructure that agent-native software needs. Across these layers, buyers need generated output to fit a workflow and survive production scrutiny.
AI tools now sit inside many standard development environments, where their value depends on workflow fit and oversight. Two university studies describe how developers work with these tools. Pace University said AI handles routine tasks and generates drafts while developers guide design, review output, and manage complexity in production systems. Developers must review and adapt generated code because they remain accountable for correctness and production behavior. Michigan Technological University said programmers who use AI to speed their work will be more valuable than those who ignore it.
Pace University’s Dr. Christelle Scharff also said:
“I believe that the big deal is that the tools related to AI are now in the hands of everybody, while earlier advances were confined and used by people in tech.”
Forecasts still point to growth, but not to an easy landing. The U.S. Bureau of Labor Statistics projects computer and information technology occupations to grow much faster than the average for all occupations through 2034, with about 317,700 openings each year. Pace’s analysis said roles requiring system-level thinking, human judgment, and accountability are less likely to be fully replaced, including architecture, security, infrastructure, platform, and technical product work.
The tension appears at the entry level. Michigan Tech said unemployment rates for new computer science and computer engineering graduates have risen compared with rates in some other fields. As routine coding automates, total demand can stay healthy while new graduates face a harder entry market.
Taken together, the data point to continued demand for AI labor alongside uneven entry-level prospects. The market is selective about people who can carry production responsibility, and Hercules’ slate is one company-level expression of that narrower truth.
Hercules is hiring people to make its agent more accurate, faster, and cheaper while keeping it production-ready. Its public criteria turn the question around: can the candidate show evidence that their own work demonstrates the same operating discipline?
Working in frontier tech? Zero G Talent tracks the openings: see every open Hercules role, browse frontier tech jobs, the companies hiring, and the people building the field.