AI Guides › Playbooks
By Nigel Guy · 8 min read
The "agent manager" post has a familiar shape: a new job title, a famous publication, a big-name employer and a six-figure salary, often "up to $260,000" (roughly £195,000 at the time of writing). Each part sounds plausible: the title is real, the employer is real, and before long the pay figure is being believed because of the company it is pinned to. That is the bad default, where one true claim vouches for three unchecked ones and someone then buys a course certificate on the strength of the lot.
The rule: check each claim on its own, against the primary source, and keep only what survives. Then prove the skill with a run you can show, not a certificate you can't.
This playbook gives you two named tools: the Claim Ledger for sorting what's real, and the Proof Run for showing you can do the work.
Write each claim in the post as its own row. Next to it, put the primary source (the publication itself, the employer's own posting) and one of four verdicts: Holds, Corrected, Unverified or Cut. A secondary blog that repeats the claim is not a source. Neither is another social post.
Here is the ledger for the agent manager claims as of 4 October 2026.
| Claim as it circulates | What the primary source shows | Verdict |
|---|---|---|
| Harvard Business Review named the role | HBR published "To Thrive in the AI Era, Companies Need Agent Managers" by Suraj Srinivasan and Vivienne Wei on 12 February 2026. It presents the agent manager as a new kind of leader who runs how AI agents learn, collaborate, perform and work safely alongside people, and compares the shift to the rise of the product manager. | Holds (as an argument in a management article, not an official job classification) |
| Salesforce has people doing it | The HBR piece opens on a Salesforce "support agent manager" who runs a fleet of AI support agents on Agentforce and starts and ends the day in dashboards and scorecards. | Holds |
| Companies are hiring for it | Salesforce has posted roles titled "Help Agent Manager, Service Cloud" and "Senior Help Agent Manager, Slack & Trailhead". The duties are resolution-rate improvement plans, scorecards, error analysis, a "book of Agent Skills" and roadmap work with product and engineering. | Holds, but under "Help Agent Manager" titles, not "AI Agent Manager" |
| It pays up to $260k (about £195k) | The Help Agent Manager, Service Cloud posting states a typical base range of $117,400–$177,600 (roughly £88,000–£133,000), rising to $140,900–$193,700 (roughly £106,000–£145,000) in selected San Francisco and New York areas. The senior Slack & Trailhead role appears on job boards at $143,400–$216,900 base (roughly £107,500–£163,000). No "agent manager" posting we found has a $260k ceiling. | Corrected |
| Any named list of other big employers "hiring for it now" | Not checked against each employer's own careers page. | Unverified, so leave it out |
| Search "AI agent manager" and you'll find thousands of jobs | ZipRecruiter's "Ai Agent Manager" page (checked 3 October 2026) reports an average of $26.51 an hour (about £20) and mixes in purchasing agent managers, call-centre sales managers and engineering roles. | Cut. That keyword search tells you nothing about this role |
All the salaries above are US base pay, quoted in dollars as posted; the pound figures are rough conversions at about £0.75 to the dollar, so check the current rate. A US band, especially a San Francisco or New York one, tells you little about UK pay, and we found no verified UK salary data for this title.
We could not tie $260k (about £195,000) to any agent manager posting. The likely culprit is a pay band from a more senior "AI" job, such as a product manager or principal-level role, that has drifted onto the catchier title. Treat any salary you can't match to a named posting as Unverified.
Once the ledger is done, the real claim left standing is this: companies want people who can set limits for agents, measure them and correct them. You can't certify that. You can only show it. The Proof Run is a single, small, documented exercise with a three-agent team, an approval gate and a scorecard.
You are helping me design and run a small, supervised AI agent team as a portfolio exercise that demonstrates agent-management skills to an employer.
Context:
- The task the team will complete: [REAL_TASK, e.g. "a 600-word briefing on X for audience Y"]
- The tool or platform I am using: [TOOL_OR_PLATFORM]
- Sources the team may use: [ALLOWED_SOURCES]
- Where the final output would go if approved: [DESTINATION, e.g. "a draft email I send myself"]
Goal: a written run plan plus a scorecard I can fill in, so that a hiring manager could see how I set limits, kept a human sign-off in place and handled a correction.
Do this in order:
1. Define three agents: a Researcher, a Drafter and a Reviewer. For each, write its job in one sentence, what it may read or change, and a short "must ask first" list (anything that sends, publishes, deletes, spends money or touches personal data).
2. Define the hand-offs: what each agent passes to the next, in what format.
3. Design one approval gate: the exact point where nothing moves on without my explicit yes, and what I must check before giving it.
4. Plan a deliberate rejection: write the specific feedback I will give on the first draft, and say what a good corrected version must change.
5. Produce a scorecard with these rows, each scored Pass / Partial / Fail with a one-line note: stayed within its limits; asked before restricted actions; approval gate held; correction applied fully; final output usable without heavy rewriting; sources traceable.
Output format: headed sections for steps 1–4, then the scorecard as a table with empty score and note columns.
Constraints:
- Do not invent capabilities my tool may not have. If a step depends on a feature, say "check your tool supports this".
- Nothing goes outside my own accounts during the exercise.
- If any bracketed input is missing or vague, ask me for it before writing the plan.
Before you answer, check: does every agent have a "must ask first" list, is there exactly one clear approval gate, and could someone else repeat this run from your plan alone?
Fill in the task, your tool, the sources you allow and where the output would go. Pick a task you can judge the quality of yourself.
Then run it, and keep the plan, the rejected draft, your feedback, the corrected version and the scorecard. That bundle is the portfolio piece. If the gate failed, keep that in with what you changed. An honest Fail with a fix shows more judgement than a clean sheet.
Priya, an operations coordinator in Leeds, wants to move into an agent-facing support role. Her ledger cuts the $260k (about £195k) figure and leaves her with one real target: help-agent and support-agent manager roles, where the job is resolution rates and scorecards. For her Proof Run she picks a task close to that work: drafting a help-centre answer to a common customer question from three public documentation pages. She rejects the first draft because it stated a refund window that wasn't in the sources. The corrected draft cites the page it took each fact from. Her scorecard marks "approval gate held" as Pass and "sources traceable" as Partial, with a note on what she'd tighten. That page goes with her applications.