Book a 20-minute call
Menu

Research · 8Hats Lab

How work changes when AI agents join the team.

8Hats Lab studies how people and AI agents exchange context, check results, and make decisions together, and what has to change in workflows, management frameworks, and company policies when AI agents become contributors. The unit of study is the whole work system, not the model. This page states what we study, what is already public, and what is still open.

For the chief executive:who owns agent-produced work, how much effort review actually takes, and which operating rules the company writes down.

For the technical leader:task boundaries, the context an AI agent works from, verification, permissions, and failure handling.

Explore the researchOpen the public records

Book a 20-minute call Prefer email? di@8hats.ai

Every record in the register keeps its publication status, permanent identifier, supported use, and limit.

Research streams

Three streams, mapped before the detail.

Each stream asks one question about work, examines one aspect of how the company operates, and carries its status with it. The map is the summary; every entry anchors to the full section, where the sources and the limits live.

  1. Frameworks under agents

    The research question
    What happens to a management framework when part of the team is an agent?
    Aspect of work examined
    How work is managed: goals, roles, cadence, evidence, and decision rights.
    Why it matters in a company
    Operating rhythms are built for human output. AI agents produce work at a different speed, so acceptance, verification, and escalation need explicit rules.
    Where policy adapts
    Decision rights, permitted tools, and escalation routes become written rules. A repeated agent error becomes a rule, a test, or a gate.
    Current status
    Working model on top of the Scrum Guide's own definitions. Not an official extension of Scrum; not validated against a controlled comparison.
    Open the stream

  2. Working with agents

    The research question
    Which parts of managing work do not transfer to an agent, however capable it becomes?
    Aspect of work examined
    Review, coordination, throughput, and release decisions.
    Why it matters in a company
    More agent output does not shrink the review queue. Someone still accepts each result and carries the residual risk.
    Where policy adapts
    Authorization and record keeping: who may release a result into use, and what record the decision leaves behind.
    Current status
    Public evidence cited within its limits, plus an 8Hats lifecycle proposal that has not been tested against alternatives.
    Open the stream

  3. How agents learn

    The research question
    How does a capability get into a person and an agent at the same time, and how do we know it arrived?
    Aspect of work examined
    Learning, assessment, deployment, and observation in real work.
    Why it matters in a company
    A program that ends at completion cannot show what people and AI agents can do afterwards. The gap stays invisible until the work depends on it.
    Where policy adapts
    Eligibility: what a demonstrated capability may be used for is an organizational decision, and certification is an input to it rather than a substitute.
    Current status
    Specification. HALA (the Human–AI Learning Architecture) and its interfaces are documented; controlled outcome validation and peer review of the dual ladder remain open.
    Open the stream

Research model

Trust = Alignment × Reliability, as a working model.

Working research model

Trust has two failure modes.An AI agent can misunderstand the reality it acts in. A human–AI system can also fail to produce a correct outcome repeatedly. The two areas below study them: alignment first, reliability second.

A working model we use to organize measurement, not an established quantitative law. Nothing on this site is scored with it, and no readiness number is derived from it.

Both factors are required and neither is sufficient: if either factor falls to zero, trust falls to zero. What the model does not yet cover is who carries the residual risk when a result is wrong. That question stays open in the second stream below.

Alignment

Does the AI agent's picture of the user and organization match the reality it acts in?

Human contribution

People make intent, constraints, corrections, and decision rights explicit.

AI agent contribution

AI agents preserve relevant context, calibrate claims, and update their knowledge after correction.

Open questionHow can that match be measured without confusing agreement with accuracy?

Reliability

Can an AI agent or a human–AI team produce correct, defensible outcomes repeatedly?

Human contribution

People set quality standards, review evidence, handle exceptions, and own release decisions.

AI agent contribution

AI agents produce traceable work, respond to failure, and behave consistently across repeated tasks.

Open questionWhich failures belong to the AI agent, the workflow, or the decision system around it?

Evidence register

Public records, with their limits attached.

Every record keeps its publication status, permanent identifier, supported use, and limit in the same row. A dataset does not lend its standing to a preprint, and neither of them stands in for a result in any one company.

  1. Public dataset

    K-Forge Evaluation Dataset

    72 questions across production courses, 7,818 model responses from ten LLMs, and 28,593 evaluative judgments.

    Zenodo · CC BY 4.0 · doi.org/10.5281/zenodo.20633464

    Supports
    Comparative evaluation of model responses to production course-design questions.
    Limit
    A bounded course-design dataset, not a universal benchmark for AI agent trust.
    Open record
  2. Public dataset

    CourseFactory Workflow-Mining Dataset (De-Identified)

    Dataset record and licensing information are available on Zenodo.

    Zenodo · CC BY 4.0 · doi.org/10.5281/zenodo.20517429

    Supports
    Inspection of de-identified records used in workflow-mining research.
    Limit
    A dataset record; it does not show that an 8Hats product changes an organizational outcome.
    Open record
  3. Public preprint

    Mining AI-Assisted Course Design Workflows at Production Scale

    A Research Square preprint on workflow mining at production scale.

    Research Square · doi.org/10.21203/rs.3.rs-10310358/v1

    Supports
    A method and findings for mining AI-assisted course-design workflows at production scale.
    Limit
    A preprint. It has not completed peer review.
    Open record

These records come from course-design and workflow-mining research. They are evidence about evaluation and workflow mining. They do not validate the Scrum proposal, HALA, or any commercial outcome.

Stream 01

Frameworks under agents

What happens to a management framework when part of the team is an agent?

Scrum is the first worked example. Its logic — transparency, inspection, adaptation — survives the arrival of agents. Its familiar implementation does not. Agents make production cheap, so the scarce resources become an unambiguous goal, an admissible account of organizational truth, a boundary on autonomy, independent evidence, and the right to make the final call. The eight objects below are where the weight moves.

A backlog item turns into an execution contract.

Expected result, inputs and their versions, permitted tools, forbidden actions, acceptance criteria, tests and control cases, evidence format, stop and escalation conditions, and the owner of the final decision.

Two loops instead of one.

A human loop runs around an agent loop that turns over in minutes: goal, priority, decision, and accountability outside; task, run, check, and correction inside. People manage the boundaries of autonomy and the exceptions, not each operation.

Eight objects, and where each one moves
Roles
In a team of peopleThe Product Owner orders the backlog, the Scrum Master supports the process, Developers do the work.
In a team with agentsThe Product Owner sets the goal, the admissible organizational truth and the decision rights. The Scrum Master designs agent envelopes, gates, escalation routes and observability. Developers write the assignments, check the results and carry the accountability. Agents execute, research, criticise and verify — and own no decision.
Cadence
In a team of peopleOne cadence. A Sprint of a month or less contains the work.
In a team with agentsTwo nested loops. A human loop — goal, priority, decision, accountability — around an agent loop that turns over in minutes: task, run, check, correction. Planning or a Daily for each agent run is pointless; people manage the boundaries of autonomy and the exceptions, not each operation.
Backlog
In a team of peopleAn ordered list of what the product needs. “Improve the sales advisor” is a workable item.
In a team with agentsAn item becomes a small execution contract: expected result, inputs and their version, admissible picture of the world, permitted tools, forbidden actions, acceptance criteria, tests and control cases, evidence format, stop and escalation conditions, and the owner of the final decision.
Daily
In a team of peopleThe team inspects progress toward the Sprint Goal and adapts the plan for the day.
In a team with agentsStatus is collected automatically, so the meeting is spent on deviations. Which runs finished. Which gates did not pass. Which statements turned out ambiguous. Which agents are waiting on a human decision. Where two agents contradict each other. How much human attention acceptance is costing. Which repeated errors should become tests or rules.
Definition of Done
In a team of peopleThe Increment meets the shared standard and can be released.
In a team with agents“The artifact exists” stops meaning “the work is done”. Done also requires: the result is reproducible; inputs and versions are fixed; tests passed; claims trace to their sources; an independent check was performed; permissions and safety held; the change can be reversed; the unknown is not masked by confidence; and the result is fitted into the shared picture of the world.
Planning
In a team of peoplePlanning selects the work and the way the team will start it.
In a team with agentsPlanning also designs parallelism: isolation of workspaces, dependencies, integration order, independent judges, checkpoints, owners of conflicts, and a limit on how much work may sit under verification at once. Parallelism is what creates the new risks — two agents editing one file, different versions of the truth, errors correlated because the prompt was shared, and hundreds of locally correct changes adding up to a globally inconsistent system.
Velocity
In a team of peopleClosed items per Sprint give a rough sense of capacity.
In a team with agentsClosed items are trivial to inflate with agents. Verified throughput is the measurable quantity: how many results were actually accepted, the gate pass rate, returns and corrections, contradictions found, human time spent on verification, the cost of one accepted result, the change in the product outcome, and the number of errors turned into permanent regression tests.
Retrospective
In a team of peopleThe team chooses changes that improve its own effectiveness.
In a team with agentsThe agent infrastructure is updated too: the picture of the world, prompts and roles, tools, test datasets, permissions, escalation routes, error examples and quality criteria. A repeated agent error stops being a verbal remark and becomes a rule, a test or a gate.

StatusEstablished: the Scrum Guide's own definitions of the framework, its accountabilities and its artifacts. A working model: everything in the right-hand column. It is an 8Hats operating proposal, not an official extension of Scrum, and it has not been validated against a controlled comparison.

A second worked example — project management — is in study.

Ask about this stream

Stream 02

Working with agents

Which parts of managing work do not transfer to an agent, however capable it becomes?

Four boundaries keep reappearing in public evidence. Each one separates a thing agents are good at from a thing they do not acquire by getting better. Read together they describe why more capability moves a bottleneck rather than removing it.

Four boundaries

  1. Capability is not self-management.

    A frontier model's own prompting guidance describes what to do when it gives too few progress updates, stops early, drops constraints during compaction, expands scope, or answers from memory instead of searching. Read as release notes this is routine. Read as a description of operating conditions, it says the most capable model in a system is not the same thing as a system capable of managing work.

    Vendor documentation describes one model family. It is evidence about operating conditions, not a measurement of them.

  2. Communication is not coordination.

    CooperBench put two agents on more than 600 collaborative coding tasks across 12 libraries and four languages, with a shared objective, a channel between them and executable tests. The paired agents performed about 30% worse on average than a single agent doing both parts. The transcripts still read as collaboration: messages were vague or badly timed, agents departed from commitments they appeared to accept, and each formed wrong expectations about the other's plan. Communication moves symbols; coordination aligns action over time.

    A coding benchmark cannot stand in for every organization or every collaboration pattern. It isolates a failure mode; it does not measure its prevalence.

  3. Activity is not productivity.

    Reuters reported that Meta's internal platforms saw 220% more changes year over year while changes that reached users as new or improved features rose 36%. The two figures have different denominators and are not two stages of one funnel — their value is comparative. Meta's own RADAR paper attributes more than 80% of the growth in code supply to agentic AI and reports that the share of changes receiving timely review fell. DORA's 2025 study describes AI as an amplifier of whatever the delivery system already is.

    Reported internal data and a company's own engineering paper are not audited productivity statistics, and none of this establishes that AI caused the incidents. What it supports is narrower: production scaled faster than the organization's ability to convert it.

  4. Production is not a verified outcome.

    An agent can run several iterations while a team is still discussing one. It cannot thereby decide which result the organization should rely on, which residual risk is acceptable, or who may release it. Accountability did not move with the speed. Calling an agent a Developer, putting its output in a backlog, or letting it say “done” does not create a Definition of Done.

One lifecycle

Six states, kept apart.

A task is not done when an AI agent stops typing. It moves through states, and each state answers a different question. Select a state to see what changes in it, what an AI agent or an automated control may do, who holds the decision right, and what evidence the workflow keeps.

Execution complete

What changes at this state
The work exists. Nothing about it has been checked yet.
What an AI agent or an automated control can do
An AI agent can execute: produce the result, fix the inputs and their versions, and link claims to their sources so the work stays traceable.
Who holds the decision right
The executor, within the scope the assignment grants. Executing does not give an agent decision rights.
Required evidence or record
The artifact, the inputs and versions used, and the assignment it answers.

When something fails, the loop writes back

Corrections and failuresTask instructionsTests and control casesWorkflow rulesCompany policy

  • Verification is not authorization. A result can pass every check and still wait for the decision to release it.
  • Automated checks and policy-approved low-risk transitions remain possible. People do not approve every operation.
  • Accountability remains with the people and institutions that hold and delegate decision rights. Automation does not acquire accountability by applying a rule.
Illustrative lifecycle with six states: Execution complete, Evidence-qualified, Verified, Authorized, Released, Outcome measured.

Illustrative scenario. One task moving through one workflow. The states are kept apart so each can be inspected on its own.

Baseline wordingexecution complete → evidence-qualified → verified or Done → authorized → released or put into use → outcome measured

The rights stay separate too. An agent or an automated control may verify defined criteria. A deterministic gate may qualify a predefined low-risk transition under a policy people approved in advance. A decision-right holder accepts residual risk and authorizes a consequential action. The person or institution that delegated those rights remains accountable. Automation does not acquire accountability by applying a rule.

Stream 03

How agents learn

How does a capability get into a person and an agent at the same time, and how do we know it arrived?

HALA — the Human–AI Learning Architecture — couples human development and agent development instead of training them separately. Its design claim is that matched capability, not maximum capability, predicts whether a human–agent pair works: seven tiers on the human side matched against seven on the agent side, a dual ladder rather than two independent scales. The mechanism is small: pair the two baselines against one shared objective, practise together with scaffolded autonomy and explicit boundaries, and refresh the working agreement when either side or the organizational firmware changes.

The chain exists so the result of learning is not a good answer in one session but a verified change in a system that keeps acting afterwards. Each object is inspectable, so a claim about capability can be traced back to the run and the assessment that produced it — and forward to what was observed in the field.

Nine traceable objects between a course and the field
  1. CoursePackage
  2. LearningRun
  3. LearningAssessment
  4. RuntimeBuild
  5. RuntimeAssessment
  6. AgentAssessment
  7. Activation
  8. Deployment
  9. FieldObservation

What this changes

For corporate learning
Completion stops being the unit of account. The unit becomes a capability that was assessed, activated, deployed and then observed in real work. A programme that cannot produce a FieldObservation has not shown that anything transferred.
For universities
A certificate can attest readiness. It cannot create authority. The eligibility gate a change must pass — legitimate, safe, verified, and reversible enough — is an organizational decision, and certification is an input to it rather than a substitute for it.
For the agent
Autonomy grows only with mutual readiness, and the boundary stays human. Inside it the agent acts; at it, a person decides.

StatusSpecification. The layered framework and its interfaces are documented; controlled outcome validation remains open, and no generalized performance claim is made for it. The dual ladder is a measurement proposal that has not completed peer review.

Ask about this stream

Publication discipline

A result keeps its object, status, and uncertainty.

  • Open the object.Permanent records and named-author work link to their source.
  • Show the status.Dataset, preprint, and under-review work do not borrow credibility from one another.
  • Keep the limit.Context evidence does not become product efficacy evidence.

Questions worth testing with others

Measurement of alignment and reliability, human correction in production workflows, and how accepted lessons update organizational practice.

Research collaboration

Research collaboration

Ask about a research stream.

Requests reach the research team directly. What we can share depends on the stream: some material is published, some is still under study, and we say which is which before a call.