An agent that works for the worker and earns autonomy in stages.
Fromplatform ranking→Toa personal agent
The agent is the worker’s representative to the network. It turns intent into a pathway, tracks what the worker can demonstrate, reads demand, and brings forward opportunities within limits the worker sets. It isn’t the network, and it shouldn’t be paid in a way that makes it work for someone else. The evidence supports agents that search, evaluate and recommend. It doesn’t yet support agents that negotiate for a worker or commit them to anything, so autonomy should be earned in stages, each with its own test.
Looking forward
Résumés look backward. The human-capital agent looks forward. [CN]
The agent works in a loop. It starts from where the person wants to go and what they can demonstrate today. It identifies what is missing, where to acquire it and how to demonstrate it, and where the person can use it to earn. Then it learns from what the work revealed and decides what comes next [CN].
That makes it a capability navigator rather than a course recommender. A course recommender optimises what to consume next. The agent optimises the movement from intent to demonstrated capability to paid work.
The agent is not the network
| Who | What it provides |
|---|---|
| The network | Participants, knowledge, training, assessment, work, evidence infrastructure, discovery, demand signals and coordination |
| The person’s agent | Interpreting intent, building the pathway, finding gaps, representing the person’s interests and supporting decisions |
| The person | The final decision on what to learn, what to disclose and which opportunities to accept |
The network wants transactions, employer access, training demand and assessment volume. The worker may want privacy, a higher rate, selective visibility, time to learn and the freedom to say no. The two sets of interests overlap, but they are not the same [CN].
The network optimises coordination. The agent optimises one person’s trajectory.
Who the agent works for
An agent’s loyalty follows its funding. If employers or the network pay it per placement, it will drift toward placements. There are four candidate models, and none is proven. The agent could be publicly funded as part of employment services, just as public employment services already fund matching tools. Workers could pay a subscription, though their willingness to pay before the network has shown any value is likely to be low. Employers could fund it under enforceable fiduciary rules and auditable logic. Or an intermediary or nonprofit could run it for workers in a capability family.
Trust also depends on who builds the agent. People judge algorithms less fair when outsiders build them, which is one reason France’s public employment service chose to develop its own tools [1]. The funding model is an open design question, and it belongs in the pilot plan.
An agent is only as loyal as the way it is paid.
Go deeper
Autonomy in stages
The concept’s long-term design is agent-to-agent. A demand agent breaks down a requirement, and the worker’s agent represents capability, availability, intent and price. The two exchange information, assess fit, negotiate terms, and bring a final opportunity to the human [CN]. The evidence supports the early steps. AI can reduce matching friction in bounded tasks. Safe autonomous negotiation is unproven, and the two sides want different things. The worker’s agent wants privacy, pay and options, while the demand agent wants speed, low cost, commitment and information.
| Stage | The agent may | The person decides | Evidence needed before the next stage |
|---|---|---|---|
| 1. Search and evaluate | Find learning, work and gaps in the person’s evidence | Everything | People act on the recommendations and save time |
| 2. Recommend | Rank and explain options within the person’s constraints | What to pursue | Explanations match what the agent actually did, and outcomes beat unassisted search |
| 3. Communicate | Contact demand agents within set disclosure limits | What is disclosed and accepted | Disclosure errors and complaints stay low |
| 4. Negotiate routine terms | Negotiate within floors and bounded authority | The final engagement | Terms are at least as good as those the person gets alone |
| 5. Contract | Not proposed yet | Not applicable | Legal clarity on agency, liability and employment status |
Autonomy is earned one stage at a time, and each stage has its own kill signal.
Keeping the human decision real
A human in the loop can quietly become a rubber stamp. People defer to automated recommendations more under heavy workloads and when tools show precise-looking scores. They defer less when tools show uncertainty, present information rather than verdicts, and leave people accountable for the decision [1]. So the agent should show its confidence and its reasons, and make the person’s decision genuinely theirs, not a click to confirm.
Seeing the whole picture
A pathway that raises pay isn’t always one that leaves someone better off. The Federal Reserve Bank of Atlanta modelled a common healthcare career ladder. Once lost public benefits were counted, a worker on that ladder could take around fifteen years to be better off than someone in a minimum-wage job. The bank’s planning tool, which shows people where those cliffs fall, has been piloted in U.S. state workforce programs [2]. An agent that recommends pathways and work should see the same picture: pay, benefits, repayment obligations (Shift 8) and the time each step takes.
The agent and the market signal
Signals of demand across the network let the agent ask which capabilities are becoming valuable and what the person is missing [CN]. They also create a risk of herding. If many agents act on the same signal, many people train for the same capability at once, and uniform algorithms can reduce match quality across a whole market [1]. The agent should present signals with their uncertainty, favour capabilities that transfer, and keep a person’s options broad rather than steering everyone toward the same peak.
Not our invention
Public employment services already use AI to profile job seekers, suggest skills and recommend jobs. Large employers use internal tools that recommend moves and learning to their staff [1]. Planning tools such as the Atlanta Fed’s already help people see the financial consequences of career steps [2]. The addition here is an agent on the worker’s side that spans formation, evidence and discovery. Its authority is bounded, and its funding is designed to keep it loyal.
What could go wrong
- A misaligned agent.If the wrong party funds the agent, it optimises for placements.
- Over-trust.People accept recommendations without scrutiny [1].
- Disclosure errors.The agent reveals more than the person intended.
- Herding.Many agents push many people into the same capability.
- Liability gaps.No one is answerable when a commitment made by the agent goes wrong.
How we’d test it
| The test | What we would measure | What would end the idea |
|---|---|---|
| Recommendation quality | Time to paid work, pay and regret for people following the agent’s recommendations, against human counselling | No gain in time to paid work or pay, or higher regret |
| Disclosure safety | Disclosure errors and complaints under tiered disclosure | Material errors, or people can’t audit what was shared |
| Bounded negotiation (later) | Terms negotiated with human oversight, against the person’s own negotiation | Worse terms or material errors |
How strong is the evidence?
- Established
- AI tools can speed search and recommendation in bounded settings [1]
- Evidence-supported
- Automation bias and rubber-stamping are real risks, with known design mitigations [1]
- Hypothesis
- A worker-side agent that spans formation and discovery improves outcomes
- Unknown
- Whether agents can negotiate safely, and which funding model keeps the agent loyal to the worker
Sources
- [1]Broecke, Artificial Intelligence and Labour Market Matching, OECD Social, Employment and Migration Working Papers (2023)Checked; the studies it summarises were not re-checked
- [2]National Governors Association, resource page on learning and employment records, including its account of the Federal Reserve Bank of Atlanta’s CLIFF toolChecked; the Atlanta Fed modelling it reports was not re-checked
- [CN]Open Commerce for Human Capital: concept document consolidated from the discussion (internal)Provided
Spotted a source we missed or a rule we misread? Tell us what we got wrong.