Detailed questions, examples, and sources for our three research streams. The unit of study is a team of people and AI agents.
Research 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.
Does the AI agent's picture of the user and organization match the reality it acts in?
People make intent, constraints, corrections, and decision rights explicit.
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?
Can an AI agent or a human–AI team produce correct, defensible outcomes repeatedly?
People set quality standards, review evidence, handle exceptions, and own release decisions.
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?
Stream 01
What happens to a management framework when part of the team is an AI agent?
Scrum is the first worked example. Its logic — transparency, inspection, adaptation — survives the arrival of AI agents. Its familiar implementation does not. AI 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.
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.
A human loop runs around an AI 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.
Stream 02
Which parts of managing work do not transfer to an AI agent, however capable it becomes?
Four boundaries keep reappearing in public evidence. Each one separates a thing AI 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
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.
CooperBench put two AI 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 AI agents performed about 30% worse on average than a single AI agent doing both parts. The transcripts still read as collaboration: messages were vague or badly timed, AI 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.
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.
An AI 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 AI agent a Developer, putting its output in a backlog, or letting it say “done” does not create a Definition of Done.
One lifecycle
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.
When something fails, the loop writes back
Corrections and failuresTask instructionsTests and control casesWorkflow rulesCompany policy
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 AI 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 does a capability get into a person and an AI agent at the same time, and how do we know it arrived?
HALA — the Human–AI Learning Architecture — couples human development and AI agent development instead of training them separately. Its design claim is that matched capability, not maximum capability, predicts whether a human–AI agent pair works. 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.
What this changes