Research on work and agents
How the maps were made
Explore the work behind a role and how its execution could change with AI and agents.
Start from the work
RecruitSignal is a theoretical research project exploring how work changes with AI and agents. The maps describe outcomes, tasks, decisions, artifacts and handoffs within a role. Context such as seniority, team structure or local policy can change a path or who decides.
Sources and interpretation
Public descriptions of work, professional guidance and technical documentation informed the maps. Sources include occupational frameworks, Google SRE guidance, W3C accessibility resources and role-specific professional material. Workflow structure and proposed AI responsibilities are research interpretations informed by those sources. Useful references accompany the work they inform.
Models within agent systems
We examine both model capabilities and the systems around them: triggers, tools and integrations, persistent state, continuing execution, feedback and verification, permissions and human checkpoints. Each connected workflow puts the work, information passed forward and decision paths together. Agent execution is conditional on the specified inputs, tools, permissions and result checks; a whole job is not automated merely because one task can be delegated. Additional responsibilities connect to the work they affect and are counted separately from changes in who executes existing work.
How to read the labels
Human-performed means a person performs the core activity; AI may support preparation or recording in the stated setup. AI-augmented means AI performs meaningful work while a person still performs a judgment, decision or action inside the task. Agent-executable means an agent can perform all in-task steps under pre-set scope and criteria, verify the resulting state and hand exceptions to the named owner. Human judgment, authority and accountability can remain in every category; these labels describe execution, not the absence of responsibility. New with AI identifies a responsibility introduced by the adopted AI or agent capability, including AI-specific work already present in the source workflow. Adding an entry during research, repeating ordinary work or splitting a task does not make it new. “Includes new” marks a new responsibility inside a broader task. Execution labels remain separate. New-work totals count distinct maintained-work types relevant to the selected scope, including shared work; domain cases and embedded facets do not become extra types. The same audit type is counted consistently whether standalone or inside a larger contract. These are coverage counts, not numbers of deployed systems, role-owned duties or profession impact. Selected details distinguish the maintainer, domain criteria owners and consuming roles; a reference home is a place to find the example. Reuse sufficient maintained cases and provider controls for the same setup and behavior. Dotted links to execution labels qualify the agent contribution; they do not make ordinary human work wait for an agent check.
Review and research reference
The work combines public-source research, dual-model review with Claude Opus 5.5 and GPT-6.1 Sol, and human review. The 28 October 2026 research revision reviewed the full existing corpus with Codex across all five roles. Independent Claude reviews challenged every workflow-level decision and the proposed changes using selected original task contracts; a final cross-role challenge checked the reconciled responsibilities and shared-use boundaries. On 29 October, task-based role and stakeholder simulations using gpt-6.1-sol, claude-opus-5-5 and gpt-6-astra compared two matched representations of ten workflows. Independent scoring and adversarial review informed specific decision and handoff corrections; the responses did not establish an overall advantage for either representation. Local implementation and interface checks continued on 29 October. Model review helps find omissions and inconsistencies; it does not establish empirical validity. Wilton’s substantive review of the published work shaped the framing, Method and agent coverage.
Most source material was captured on 22–23 October 2026; specific references were refreshed where needed during the 28 October 2026 corpus revision. Source capture dates and specific capability conditions remain distinct from that review date; this pass did not refetch every reference. Counts describe coverage rather than measured productivity or employment forecasts.
Anthropic: Building effective agents · OpenAI: Agents
Why map the work?
RecruitSignal makes changes in work concrete enough to examine: how responsibilities connect, what an agent could execute and which decisions people retain. Explore a role map or try an exercise to consider how the work could change in your own setting.