AI Hiring Pulse — Week of 2026-06-22

📊 Pattern of the week — the frontier diverges

Across 677 new AI roles at 129 companies this window, the index contracted again: 677 opened, 1,478 closed, net −801. That is worse than last week's −695 and the second consecutive week of deepening contraction. The active index stands at 7,681 roles.

But one sector broke the other way. AI frontier labs posted 63 new AI roles, up from 26 last week — a 142% increase that lifted their share of total hiring from 3.7% to 9.3%. No other sector grew. Bigtech dropped from 279 to 229. Enterprise cratered from 115 to 76. Fintech halved from 25 to 13. The frontier labs are hiring into the contraction.

The surge has three distinct signatures. Each tells a different story about where the model companies think the next bottleneck is.

🧪 OpenAI's Agent Post-Training taxonomy

OpenAI posted 16 roles this week. Eight of them — half — are Post-train, and seven carry the title prefix "Agent Post-Training" followed by a sub-specialty:

• Computer Use Research — training models to operate GUIs
• Connectors Research — integrating with external tools and APIs
• Artifacts Research — generating persistent objects (code, documents)
• Context Research — managing long agent working memory
• Personality — behavioral tuning for agent personas
• Frontier Evals and Environments Research — evaluating agent capability boundaries
• API & Power Users — optimizing for developer-facing workflows

This is not "train a better base model and agents improve for free." This is a post-training pipeline per agent modality. OpenAI is treating agent capabilities as distinct training problems, each with its own research team and its own evaluation surface. The eighth Post-train role is a "Model Designer" — suggesting a meta-layer that orchestrates across the sub-specialties.

The rest of OpenAI's week reinforces the pattern: three Agent roles (including "AI Deployment Engineer, Cyber" and "Product Manager, ChatGPT Sites"), one Eval ("Safety Transparency Editor"), and four Ship including "Prototyping Lab Technician, Robotics" and "Technical Deployment Lead — UAE." The full pipeline from post-training research through physical deployment.

🚀 Mistral's full-stack applied push

Mistral AI went from 2 new roles last week to 15 this week — the biggest net positive company in the entire index at +9. Its titles are overwhelmingly "Applied AI Engineer" variants, many prefixed "NEW" — literally new positions being created.

What makes Mistral's burst distinctive is the stage spread. Most companies concentrate at one or two stations. Mistral is hiring simultaneously across four:

Ten of 15 roles carry the "applied" archetype. Mistral is not just training models — it is building the applied-AI org that turns them into products and deploys them at customers. That is the profile of a model lab becoming an application company, the same trajectory OpenAI followed 18 months earlier.

⚖️ Harvey and the vertical-agent signal

Harvey went from zero new roles last week to 8 this week, now carrying 33 active AI roles. Seven of the eight sit at the Agent station. The titles: multiple "AI Automation Engineer, Customer Education" variants and "GTM Technology Product Owner" — roles that embed agents directly into customer-facing legal workflows.

The one non-Agent hire is notable: "Research Engineer, Post-Training." Harvey is not just deploying foundation models into legal workflows — it is customizing model behavior for domain-specific accuracy with its own post-training pipeline. For a vertical AI company, that is a significant capability investment.

Harvey's pattern — 94% Agent-layer, one Post-train hire to customize — is the mirror image of OpenAI's profile. OpenAI builds the model and the post-training tooling. Harvey deploys the agent and pulls just enough post-training in-house to make it work in a domain where generic behavior is insufficient. The question for every vertical AI company is where on that spectrum they land.

📉 The enterprise freeze

The contrast with the rest of the index sharpens the frontier signal. The zero-open cohort grew this week:

Salesforce — last week's agent-layer poster child — dropped from 19 new roles to 4, net −22. The enterprise SaaS tier that was hiring aggressively to deploy agents on foundation models is pulling back. The companies that make the foundation models are not.

The structural read: the upstream half of the stack is still hiring. The downstream half — the companies that rent models and wrap them in product — is contracting. The frontier labs posted as many upstream (Pretrain + Post-train) roles this week as all of bigtech combined (17 vs 17), with one-quarter the total headcount.

🔮 Prediction watch

Grading last issue's predictions:

• 🟢 JPMorgan ≥250 active, ≥60% Agent. Current: 269 active AI roles, 183 at Agent = 68%. Both thresholds comfortably met. Two more pulses to go. On track. (Cashable 2026-07-20)
• ❌ Pharma data sovereignty — ≥2 of 3 net positive. Merck +2, Johnson & Johnson −16, GE Healthcare −1. Only Merck is positive. One more pulse remains, but J&J's −16 is decisive. At risk — likely miss. (Cashable 2026-07-06)

New predictions:

• The frontier share stays elevated. Prediction: over the next two pulses, frontier labs (ai_frontier sector) post ≥7% of total new AI roles per week. This week's 9.3% may include a one-time Mistral burst, but OpenAI's Agent Post-Training build-out and Harvey's ramp suggest structural demand. Cashable 2026-07-13. If frontier share falls back below 5%, this week was a hiring-cycle coincidence.
• OpenAI's Agent Post-Training team is structural. Prediction: OpenAI posts at least 3 new Post-train roles in each of the next two pulses. If the sub-specialty pattern (Computer Use, Connectors, Context, etc.) keeps expanding, it confirms that agent post-training is a permanent research org, not a one-time hiring push. Cashable 2026-07-13.

📰 Reading list

1. OpenAI in our index — read the seven Agent Post-Training titles as a taxonomy of what agents need beyond the base model. Computer Use, Connectors, Artifacts, Context, Personality, Frontier Evals, API. Each is a research bet.
2. Mistral AI in our index — the full-stack applied push from Pretrain through Ship in a single week. Compare against OpenAI's Post-train concentration to see two different lab-to-product strategies.
3. Harvey in our index — the vertical-agent archetype: 94% Agent-layer, one Post-train hire. The minimum viable model-layer investment for a domain-specific AI company.
4. AI Hiring by Stage — the frontier is the only sector still hiring upstream. Everyone else is at Agent or below.