Code generation cost collapsed. Verification cost did not. The engineer who got 10x more productive is the one who treats the codebase as a context substrate the model writes against, not a thing to author by hand. Below: open projects you can take on with us, the physics of what is changing, and roles we are hiring against.
Real, paid 1-3 week engagements with the Product.ai team. Each one is a problem we are working on at the frontier of engineering — and the kind of work we hire against. Pick one that pulls at you and apply.
Last updated
In May 2026, software engineering at frontier-AI firms (Anthropic, Cursor, Vercel, Linear, Stripe, Replit, Sourcegraph) is structured around one observable rupture: the LLM writes the first draft of the code. The 10x engineer is no longer the fastest typist or the cleanest architect — they are the operator who shapes the context the model reads against, designs the evaluation that defines "correct," and verifies the output before it merges. Authoring is no longer the moat. Architecting the system that the model authors inside — context, contracts, evaluation, deterministic gates — is the new craft. Senior practitioners are not threatened by AI. They are explicitly repositioning value above the keyboard tier.
The fundamental physics shift: the codebase is now a context substrate the model writes against, not a thing humans type. CLAUDE.md files, registered skills, deterministic guards, schemas, and contracts have become first-class engineering artifacts — they define what the model can and cannot do, and they govern code quality at every keystroke. Anthropic's Claude Code, Cursor's rules system, Vercel's v0 prompts, and Sourcegraph's Cody contexts all converge on the same pattern: the engineer who wins is the one who curates the model's working set.
This is why "context engineering" rose as a distinct skillset in 2025-2026. The engineer treats the repo as a system whose primary user is an LLM, not a human reviewer. They invest in the loadout — what loads when, what gets injected, what gets gated. The output is code, but the craft is upstream of the code. Engineers who skip this layer ship faster than non-AI peers but produce code their teammates can't maintain. Engineers who master it ship 10x with code that survives.
Test-driven development inverted. The 2010s pattern — write the test, watch it fail, write the code, watch it pass — assumed humans authored both. In 2026, the model authors the code; the human authors the eval. The evaluation is the specification. The engineer who can describe "correct" precisely enough that an LLM can verify it has built the moat. Anthropic's 2026 Claude Code postmortems repeatedly cite the same failure mode: comprehensive eval frameworks that systematically discovered problems weeks after users did — eval theater, where green dashboards masked degraded behavior.
The senior engineer who calibrates which evaluations produce false confidence is higher signal than the senior engineer who built the dashboard. Behavioral verification — production traffic mirroring, regression corpora, paraphrase-resistant test cases — beats stated-trust survey-style eval scores. Engineers fluent in this distinction can name which of their evals have failed silently and what they did to fix it. Engineers who can't, are still authoring under the old paradigm.
Two engineering roles that look similar from the outside have diverged structurally. The AI-Native Software Engineer uses the model — they ship product code 10x faster by treating Claude Code or Cursor as a leverage instrument. The AI Engineer / ML Engineer builds the model surface — fine-tunes, evaluates, deploys, monitors LLM systems. Both titles exist at every frontier firm; they share tooling but not craft. Conflating them in hiring produces miscast hires on both sides.
The taxonomy split is observable in compensation, in tooling, and in success patterns. AI-Native SWE roles cluster at $300-500K total comp with strong product-shipping signals. AI Engineer roles cluster at $400-700K with strong eval-architecture and inference-economics signals. Cross-mapping fails: the AI-Native engineer who tries to fine-tune produces under-evaluated systems; the ML Engineer who tries to ship product features under-invests in user-visible craft. Hire to the taxonomy.
The strongest engineering filter in 2026 is commits to a system the candidate primarily owns, with an LLM in the toolchain visible in the history. A GitHub graph that shows context files, eval suites, prompt iterations, and production fixes beats a polished resume every time. AI tools collapsed shipping cost; non-shipping reads as uninterested. Engineers who can name which prompts they iterated on, which evals they wrote, and which deterministic gates they built signal the new craft directly.
The next-strongest filter is articulating tradeoffs in plain language — a senior engineer who can explain why they chose a deterministic gate over an LLM-judge for a specific check, or why they accepted a 95th-percentile latency hit for a verification step, beats one who recites best practices. The Trust Paradox applies here: stated proficiency over-states actual fluency. Behavior beats credentials. The cleanest signal is a working system in production with the candidate's name on the commits.
The "AI replaces engineers" thesis was structurally wrong, and the senior engineering counter-narrative is now the load-bearing signal. AI didn't reduce engineering headcount at frontier firms — it raised the bar for what 'engineering' means. Boilerplate authoring collapsed; system-level design, eval architecture, and deterministic-gate engineering expanded. Firms that fired their senior engineers and tried to replace them with junior + AI produced unmaintainable codebases and walked them back within 18 months.
The companion mistake: confusing velocity with quality. Devin-style "autonomous engineer" demos in 2024-2025 produced impressive task-completion metrics that fell apart under closed-loop verification — the agents marked their own work correct (Meta-Axiom C: external truth anchors are required). The senior engineer who insists on out-of-band verification (humans, deterministic checks, regression corpora) beats the engineer who accepts the agent's self-report. The frontier is in the dissent, not the acceptance, of AI-claimed completeness.
Most of our best people came through projects, not interviews. If a project pulls at you and the trial goes well, the role conversation follows.
Twelve-minute Hireflix video, async. Then a 30-60 minute chemistry call. Then a paid 1-3 week project alongside the team. We will know within a week whether to move forward.