Agents are holding named jobs now, and the numbers are starting to show it
Six job types, 38 percent of merged pull requests at one company, and 20 hours a week of sales admin reclaimed — the agent workforce now has a job description.
- Enterprise AI agents perform six primary jobs — answer, draft, check, classify, monitor, and act — and require four foundational elements: approved data, user-specific permissions, defined rules, and audit logs. m:…7852
- At Runway, cloud agents drove 38% of merged pull requests in October 2026, changes approved by AI agents were reverted at a rate of 0.1% compared with 0.5% for human-approved changes, and the median wait time for a first pull request approval fell from over two hours in September to five minutes by mid-November. Self-reported m:…7556
- One account executive reclaimed over 20 hours per week by automating administrative tasks, with post-call follow-up time reduced from 25 minutes to 3 minutes, and reps using three or more AI tools alongside calls book 2.4 times more meetings. Self-reported m:…1822
- A McKinsey survey from August 2026 found 40% of large organizations were scaling AI agents in at least one function, while a Deloitte Tech Trends 2026 report indicates 11% of organizations have agents in production, with 38% piloting them. m:…7852
Six jobs
Enterprise AI agents perform six primary jobs: answer, draft, check, classify, monitor, and act. AI agents work through tasks in multiple steps using language models, company data, and tools — the same ingredients in every deployment, with different jobs built on top. m:…7852
Each job is specific. An answer agent responds to questions using only approved sources and cites them. A draft agent builds a first version of a document from approved material for human review. A check agent compares work against a reference to identify missing, duplicated, or unbilled items. A classify agent labels incoming documents against a taxonomy and routes uncertain items to humans. m:…7852
The last two jobs are the ones that change headcount math. A monitor agent watches sources and flags changes or risks based on predefined thresholds, and an act agent carries requests across systems to completion, such as turning an email into a quote. Enterprise agents require four foundational elements: approved data, user-specific permissions, defined rules, and audit logs. Starting with answer or draft jobs is recommended because they build the library of approved content for later use. m:…7852
The engineering floor
Runway Engineering has published the clearest numbers yet for agents doing engineering work. Its cloud agents drove 38% of merged pull requests in October 2026. The median wait time for a first pull request approval fell from over two hours in September to five minutes by mid-November. And changes approved by AI agents were reverted at a rate of 0.1%, compared with 0.5% for human-approved changes — the agents, on this measure, were more careful than the humans. m:…7556
The system behind it is a software factory, not a chatbot. Runway implemented a risk-based approval system where AI agents rate pull requests as low, medium, or high risk. Its coordinator service dispatched over 14,000 agent tasks in the first two weeks of operation, and its automated QA loop found 86 confirmed bugs in its first four days of operation. m:…7556
The monitoring loop is the part most teams skip. Runway's automated main-branch monitoring system fixed 139 out of 140 alerts without human intervention in its first two weeks, and its AI reviewer resolved or led to code changes in 93% of comment threads in the first half of November 2026. The engineering role has shifted toward identifying friction points, removing infrastructure blockers, and defining agent-friendly patterns. The long-term goal is a system that takes a description of an outcome and ships it, with human intervention only at defined points. m:…7556
The sales floor
The sales version of the same story comes from an account executive who automated their own job, mostly. The author reclaimed over 20 hours per week by automating administrative tasks, and post-call follow-up time was reduced from 25 minutes to 3 minutes. The author maintains a human-in-the-loop guardrail system for customer-facing work, and Amplemarket's MCP serves as the execution layer for the author's AI agent. m:…1822
The numbers suggest this is not one person's experiment. The number of reps using connected AI assistants grew 7 times in five months. More than 1 in 3 new prospects now receive AI-assisted outreach, and 3 in 10 new outbound opportunities now trace back to AI. Reps using three or more AI tools alongside calls book 2.4 times more meetings. m:…1822
The multi-threading numbers are the ones a sales leader will notice. Accounts reached through three or more people had 3.6 times the meeting rate of single-contact accounts, and accounts reached through six to ten people had 4.5 times the meeting rate. That is the kind of work — reaching more people per account — that an agent can do at a scale a single rep cannot. m:…1822
The inbox
Superhuman's new Partner Agents bring the same pattern into email. Superhuman Go has introduced partner agents to integrate external tools and workflows into the email drafting process. The HubSpot Context agent surfaces CRM summaries within the interface, the Gong Account Insights agent finds and summarizes recent account calls, and the Apollo Inbox Assistant agent combines email threads, meeting context, and contact history to suggest next actions. m:…0311
The checking job shows up here too. The Fact Checker agent flags outdated or incorrect product and policy claims in drafts, and checks public factual claims against web evidence. The Presentation Assistant agent surfaces matching existing decks for linking, and the Chili Piper Scheduler agent auto-drafts emails with available times and booking links. Superhuman also offers a solution partner program to build custom agents for enterprise-specific workflows. m:…0311
What to watch
The scale numbers are still modest, which is the honest context. A McKinsey survey from August 2026 found 40% of large organizations were scaling AI agents in at least one function, and a Deloitte Tech Trends 2026 report indicates 11% of organizations have agents in production, with 38% piloting them. The gap between scaling in one function and running in production is where the next year will be decided. m:…7852
The pattern across all four sources is consistent: agents do best where the job is named, the data is approved, and a human checks the output. Runway's agents were reverted less often than humans, and the sales agent kept a human in the loop for customer-facing work. The job description is written. The open question is how many organisations can meet the four foundations it requires. m:…7556 m:…1822
What is still unresolved
- Whether agent-approved code stays as reliable as human-approved code as agents take a larger share of merges
- Whether the 20-hours-a-week sales admin saving holds across teams or was one rep's stack
- How many of the 38% piloting organisations reach production in the next year
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.