Understand how AI systems behave when people put them to work. Goodfoot explores how agents gather information, use context, and judge results, drawing on recent research to explain the capabilities, limitations, and questions that matter.
Get essays that follow a question through the evidence. We unpack what researchers tested, how they reached their conclusions, and what remains uncertain, with concrete examples and source links so you can explore further.
Should an AI explain a faulty database query, test your suspicion or rewrite it? A SQL-debugging study follows people making those choices, while a simple counterexample reveals how each choice changes the checking work.
An AI can sort grant descriptions and add the money, but which mistakes deserve a closer look? A new study shows why checking the total leaves a consequential judgment about what belongs.
AI can explain a county's water-stress score, but what does it tell you about a proposed data center? A small study explores the different kinds of evidence needed to prepare a local recommendation.
An AI evaluation can change because the software changed or because people revised what counts as useful. Keeping old judgments comparable matters, but so does recognizing when the job itself needs a better definition.
What should an AI browser assistant finish for you? A browser experiment shows the promise of smaller operations, while research and shopping workflows reveal why handing over the whole task is a different choice.
When you ask AI to put a favorite tool inside another app, what else comes with it? Choosing the operation is only part of deciding how the new control should fit your work.
AI can help explain a difficult experience, but clearer wording is not the only goal. Research on chronic illness offers a way to think about what a rewrite may change, and who gets to decide.
An AI can sort your notes while leaving an important question unresolved: are these the groups your project needs? A small field study explores what it means to revise the organization itself.
What changes when a chatbot's hidden goal becomes visible during a conversation? An advice experiment suggests a useful role for disclosure, while leaving a harder question: how should ordinary systems justify those warnings?
A research paper can describe an experiment without settling every instruction needed to run it. A new tool makes some missing choices visible, while leaving researchers responsible for what the working reconstruction actually means.
Following an AI answer’s links can reveal whether it represents a source fairly. New research on planted web pages raises a further question: what evidence supports the claim on the other side of that link?
A coding agent may make a neglected investigation worth attempting. But timing the jobs already chosen leaves another question open: did the new work improve a decision enough to justify what it displaced?
An AI can explain a calculation without showing what other reasonable choices would change. A study of analytical review finds that visible alternatives weaken one result's pull, while giving the reviewer more work.
An AI answer can appear in a shared chat without becoming a shared topic. Two student studies explore what changes when partners can see the same material, and what they still need to discuss.
An AI tutor can keep a writer discussing ideas without showing that the essay improved. A classroom study explores the space between useful questions and actual revisions, where the writer still has decisions to make.
An AI can fill a grocery cart without settling whether the shopping suits your household. OpenAI’s LifeVal benchmark turns personal requirements into testable criteria, exposing both progress and the limits of a single score.
An old prompt instruction may have outlived the problem it solved. When changing AI models, deciding what to remove means asking what successful work looks like without it, and what still deserves checking.
An AI-assisted patch can be ready before its contributor is allowed to submit it. Contribution rules reveal different choices about review capacity, trust and who gets a chance to join a software project.
When AI narrates a slide deck, the words may be easy to hear while the relevant detail is hard to find. What should you examine before sending someone else the finished video to watch?
Would a crowd bet that you'll finish on time? Putting a personal goal on a prediction market can offer an outside estimate, but its confidence deserves a closer look before it becomes reassurance.
What should an AI assistant help you do before a difficult discussion? You could ask it to polish an opening statement—or use the private conversation to examine the reasons you will bring to others.
How much context should you give an AI assistant when a conversation turns personal? Useful help, identifying details and a saved record deserve separate decisions—even when the reply feels thoughtful and understanding.
Before an AI can check a program's behavior, it must decide how to set up the test. A document library's naming check reveals why that choice deserves attention alongside the answer being checked.
An AI assistant can get moving again without finishing the job. What should count as recovery: returning to an earlier screen, making the next step possible, or delivering the result the person actually needs?
An AI assistant can read every sentence and still miss a reviewer's selection. Highlighted documents reveal why carrying that choice into a revision is different from asking a model what it considers important.
Before you hand work to an AI agent, whose experience should guide you? A complaint can raise an important concern while leaving you to decide what it means for the task you have in mind.
When you ask AI to rehearse a conversation, what should become easier afterward? Choosing a precise goal changes what you practise, what help you keep and what would count as progress beyond the session.
Saving a coding agent’s command permission can keep work moving. But what should that permission cover tomorrow, when the rule may outlast the task and the reason that made the command useful today?
You can spot a problem in an AI's work and still be unsure what your feedback changed. A study of university tutors highlights the decisions between leaving a review and finding out whether it helped.
Changing an AI answer can preserve its best idea. New research asks whether models can predict those decisions, revealing a practical challenge for assistants that need to learn from what people do after chatting.
An AI agent can finish a task while leaving its reader unsure what its report means. New experiments and developers' workarounds explore why a useful explanation may require going back to the work.
Having hardware and a model file doesn't guarantee a usable AI service. New experiments show agents handling the work in between, and why that ability matters before a system can sustain another copy.
A faster model can choose an agent’s next move without doing the whole job. Tests of household and phone tasks reveal which work stays around those choices, and how the integration changes the result.
ChatGPT can turn an answer into something you adjust. A web investigation helps explain the difference between exploring those choices and asking for a new response, while the underlying information still needs checking.
Simulated participants could help you compare how an AI interviewer asks and follows up. A new study finds encouraging agreement with human interviews, while the people answering still have judgments a rehearsal cannot supply.
A copied agent skill can fit your project better and drift from its source. New research on public repositories makes the update question more specific: which changes should reach the version your agent uses?
An experimenter reports a booking reference changing in a voice agent's tool call despite repeated corrections. A new financial-query benchmark shows why particular names and numbers deserve attention alongside a recognizer's overall transcription accuracy.
Saving an explanation can help with your next AI task. A small simulation study found fewer professionals permitted AI reuse than human reuse, raising a question about who else may use the guidance you supply.
An AI tutor can find a useful question inside an incomplete answer. A field study traces that exchange alongside a harder result: fewer students assigned dialogue finished, despite favorable experiences among those who did.
A learner asks AI to write down Dutch examples, while others question a diagram or return to a list. Their next requests reveal what it takes to keep working with a spoken answer.
You can tell an assistant what you want for dinner. A new simulation tests whether that choice survives when several people's assistants have to turn their separate requests into one shared food order.
A relevant paper might offer a method to borrow, a useful comparison, or a limitation worth fixing. A new research benchmark asks whether AI search can find the help a particular project needs.
An AI answer can arrive before you choose a news article. A Washington Post test found a wider mix of topics along that search path, leaving a harder question: did readers find what they wanted?
An AI writing tool held back ready-made prose while people developed an argument. They later judged supplied passages faster, raising a practical question about what writers need to understand before asking for a draft.
Checking on a computing job seems harmless. But someone has to answer each request, and new observations at Harvard show why an AI agent's appetite for updates deserves a more careful look.
A quick AI summary can help you prepare to discuss employee feedback. But which comments shaped its headings, and what might you miss if those headings become the whole agenda for the conversation?
A simulated customer volunteers information, and an AI support agent completes the task. That helpful exchange raises a harder question: how much can a success score tell us when the customer changes the test?
How can a summary of public facts reveal a secret? In a simulated security exercise, AI agents learned to send messages that passed inspection but meant something more to a receiver that remembered earlier exchanges.
An AI agent finds a failing test, but does the software need fixing or is the test asking for the wrong thing? Two bug reports show why that question matters before writing a patch.
A finished engineering model may not tell the next engineer how its parts depend on each other. Two assignments for one small assembly reveal what a useful handoff needs to carry beyond the shape.
AI models could recognize a DNA pattern when shown it, yet often missed it when left to investigate. A biology study asks a practical question: what good is evidence an agent never looks at?
A precise specification helped an agent fix small bugs, then fell short on larger repository changes. What gets left out when the contract describes the required behavior but the job reaches across several files?
A memory system finds relevant history, yet the coding agent's repair gets worse. Selected benchmark cases show how an old record can carry useful clues and bad advice together, leaving the next worker to separate them.
Training AI reviewers on another model's reviews narrowed their scores and language in one experiment. The important question remains open: does that sameness cost a review panel the valid criticism only another reviewer would catch?
In a simulated store, higher reasoning weakened one shopping cue and strengthened another on average. But individual models broke that pattern, raising a harder question: which signals should an agent treat as evidence?
A coding agent's rejected attempts may still help decide where to spend its next round of effort. Replaying recorded searches offers a cheaper testing ground, but can yesterday's lucky branch guide tomorrow's work?
Researchers made a small benchmark reward unsafe code, then removed the poisoned tasks. In some self-improving research agents, the bad rule survived, exposing a harder recovery problem than simply replacing the test that taught it.
In a simulated room, two agents get the same unfinished task and two moves left, but only one finishes. The comparison separates getting into a useful position from knowing what to do once there.
An agent's answer can keep improving even when a fresh attempt would make better use of the remaining budget. A fixed-budget experiment explores when to continue, start again, or split the work.
A payment can succeed while its confirmation disappears, leaving an AI agent unsure whether to try again. The gap between what happened and what the agent knows can turn a sensible retry into a costly mistake.
In a synthetic shopping test, an agent chose an allowed product for a reason planted by an attacker. The transaction could satisfy its checks while leaving a harder question unanswered: why trust the choice?
A model cannot choose a tool that never starts. A one-shot server study exposes the gap between accurate function calls and reliable work, where setup failures and recovery can decide whether anything gets done.
When many agents share busy computers, starting every ready step can make entire jobs finish later. Replaying recorded work shows why the order matters, and why making agents wait is only part of the answer.
In one agent experiment, changing the system after every failure made it worse than leaving it alone. The result raises a question a failed run cannot answer: how far should its repair reach?
Showing an AI reviewer another agent’s diagnosis can steer it toward the same mistake. Experiments with simulated work raise a question for anyone seeking a second opinion: when should the reviewer see the first?
A model can write more correct tests while choosing inputs that expose fewer known faults. Controlled experiments reveal why a better score may hide an easier exam, and what happens when tests steer repairs.
Two model runs can share every visible word and still take different paths. Experiments with open models trace the difference to hidden server state, raising a practical question: what does replay need to preserve?
In games with persistent agent notebooks, replacing one teammate made the one who stayed do much of the extra coordinating. What happens when useful memory depends on a partner who is no longer there?
An old agent log leaves researchers with fragments of a vanished workspace. Filling the gaps creates a useful substitute, but how much of its value comes from the log, and how much from invention?
An agent can receive a changed requirement and still follow its old plan. In staged workflows, researchers expose a gap between knowing the latest facts and checking whether earlier decisions still hold before acting.
A discarded AI patch may expose a missed requirement or help someone discover a new one. Counting deleted lines cannot tell those stories apart, yet the difference changes what a team should fix next.
In one experiment, the AI team with the fewest connections used the most tokens. Cutting communication can save money, but what happens when a missing link also removes information another agent needs to work?
In one math experiment, the checker reported little trouble while the model grew worse. Following which answers reached the teacher reveals how a system can miss the very mistakes it needs to learn from.
An AI model kept the correct zip code but put it in the wrong field. Controlled experiments reveal why finding the right pieces and fitting them together can require different kinds of help.
Adding PDF import meant changing a Python script and the instructions an agent reads. That paired update opens a maintenance question: how do you keep one capability consistent when two different readers depend on it?
Automated researchers given expert starting ideas did not finish ahead in one study. They could test those ideas and abandon them, shifting attention toward the part of the process they could not discard: judging success.
The code survives when a new model takes over, but should every earlier guess come with it? Coding-agent studies reveal why the same history can save costly rediscovery or make a successor less effective.
An AI agent can remember an instruction while losing the warning that said it was unsafe to follow. What else must survive when past work becomes a summary, a saved note, or tomorrow’s plan?
Human competitors often returned to earlier code; the tested agents almost never did. Yet one agent recovered from setbacks without rewinding, raising a question about what counts as using history rather than simply revisiting it.
Two agents can finish their assignments and still leave work that will not fit together. Browser sessions, files, and outside changes make the return journey a design problem before parallel work can pay off.
Researchers improved an AI’s answers by changing the map to its documents, leaving the model alone. The experiment raises a practical question: when does yesterday’s shortcut help with tomorrow’s work, and when does it fail?
Work dashboards show approvals more readily than the checking behind them. A simulation and human adoption studies raise an untested workplace question: could those visible traces change whether colleagues verify the next AI answer?
Researchers added copies and paraphrases to an agent's memory without changing the correct answer, yet a voting system often changed its mind. The experiment asks when another agreeing record adds evidence rather than another echo.
A frozen model seemed to learn six math problems and forget nine without any training. Before celebrating a gain or diagnosing a setback, the experiment asks how much change the measuring process creates itself.
In a household simulation, an agent confidently repeated commands that went nowhere. Replaying the alternatives exposes a question ordinary step scores can miss: would changing that decision have made the final result any better?
A model that wrote a strong client brief slipped in the rankings when another model had to follow its advice. The experiment asks how much ability survives a handoff that allows only process guidance.
Recording every click can show an AI agent how someone worked once. Teaching it to handle a missing receipt or a changed requirement raises another question: where can it practice, and who judges success?
A claim can arrive intact while losing the evidence that made it trustworthy. Research on agent handoffs asks when an extra check can prevent that loss before other work starts treating the claim as fact.
A web agent remembers advice it cannot use, then brings it to later tasks. Experiments with persistent feedback ask what a correction should change, and how to catch a lesson that deserves to be forgotten.
When an AI coding session ends, the patch may survive while its reasons disappear. Git offers a familiar place to keep plans, test results, and decisions available to the next person or agent.
The agent finishes its patch in minutes, but the review can wait for days. Goodfoot's field report explores what reviewers need to see before a fast draft becomes code they can confidently approve.
An API field changes in TypeScript, while its Python client still expects the old name. This git-span demonstration follows the connection that ordinary import checks miss, showing how to surface it before work moves on.