How it works · TextData

Elements

Agents act, tools perform operations, workflows coordinate execution, and resources preserve what subsequent work can draw on. Explore each element and how it connects to the others.

AI agents

Determine and perform actions

In TextData, agents use large language models (LLMs) to select and carry out steps toward researcher-defined goals. An agent is an AI system equipped with tools that allow it to take actions (Anthropic 2026a).

Agent labels denote roles. The coordinator assigns work and brings results together. Task agents carry out research assignments. Research agents investigate methods and capabilities within the deep-research engine. Development agents build and maintain that engine.

Agents consult resources and call tools within workflows. The human researcher defines the inquiry and retains authority over methods and conclusions.

AI research systems and agent design

Tools

Execute particular operations

Tools are programs that perform specified operations: retrieving records, preparing texts, locating passages, or constructing measures (Model Context Protocol 2025b).

A tool returns an output that an agent can use in choosing its next action. A quotation checker, for example, can test whether words occur in a source; it cannot establish whether an interpretation of those words is sound.

Tools provide operations. Workflows arrange those operations, agents use their outputs, and reusable implementations can be retained as resources.

Quotation checks and annotation Data acquisition

Workflows

Coordinate execution

Workflows coordinate agent actions and tool operations, specifying their sequence and the conditions for proceeding. They can prescribe steps while leaving particular decisions to agents (Anthropic 2024).

Within a workflow, task agents consult resources, select and call tools, and use the outputs to guide their next actions. The workflow also specifies how results pass between agents.

A workflow organizes the work; it does not replace the researcher’s judgment about evidence or authorize a claim on its own.

Grounded Theory workflow Gioia Methodology

Resources

Preserve materials for subsequent work

Resources provide the materials on which researchers and agents draw: source texts, accumulated findings, methodological guidance, and packaged toolkit components retained for revision and reuse (Model Context Protocol 2025a).

These components include agent skills, which combine instructions, reference materials, and supporting tools for particular research tasks (Zhang et al. 2025).

Resources inform the researcher and agents. Findings and methodological decisions from an inquiry become resources for the next one. Tools and workflows describe capabilities; resources preserve materials and implementations for reuse.

Research foundation Reusable toolkits

Select any part of the drawing to learn more.

What the figure shows

The human researcher defines the question and scope, sends goals to the coordinator, and assesses the returned findings. Authority over methods and conclusions stays with the researcher.
The coordinator translates the goal into assignments for task agents and integrates their results. Workflows specify the sequence, conditions for proceeding, and how results pass between agents.
Task agents consult resources, select and call tools, and use tool outputs to guide their next actions within workflows.
The researcher directs development agents, which build and maintain the deep-research engine.
Within the engine, research agents investigate methods and capabilities. Under researcher direction, they assess components, adapt suitable tools, or develop and test an implementation.
The engine supports building and revising tools, workflows, and resources. Evidence, findings, and methodological decisions are retained as resources that inform subsequent inquiry.

Read the full diagram

The human researcher defines the question and scope, sends goals to the coordinator, and assesses the returned findings. Authority over methods and conclusions stays with the researcher.
The coordinator translates the goal into assignments for task agents and integrates their results. Workflows specify the sequence, conditions for proceeding, and how results pass between agents.
Task agents consult resources, select and call tools, and use tool outputs to guide their next actions within workflows.
The researcher directs development agents, which build and maintain the deep-research engine.
Within the engine, research agents investigate methods and capabilities. Under researcher direction, they assess components, adapt suitable tools, or develop and test an implementation.
The engine supports building and revising tools, workflows, and resources. Evidence, findings, and methodological decisions are retained as resources that inform subsequent inquiry.

Human researcher

The researcher begins by defining the question and scope. These objectives remain the basis for configuring the analytical work.

The researcher assesses the evidence, judges which claims it supports, and directs any further work. The question, analysis, or instruments can be revised as the inquiry develops.

The human researcher retains authority over methods and conclusions.

Coordinator agent

The coordinator translates the research goal into assignments for task agents, drawing on goal-directed research systems.

The coordinator connects the contributions and returns integrated findings for the researcher to examine.

Workflows

Workflows coordinate agent actions and tool operations, specifying their sequence and the conditions for proceeding. They can prescribe steps while leaving particular decisions to agents.

A workflow organizes the work; it does not replace the researcher’s judgment about evidence or authorize a claim on its own.

Task agent

Within workflows, task agents consult resources, select and call tools, and use the outputs to guide their next actions. This follows the reasoning–action approach of ReAct and programmable agent interactions.

The task-agent role differs from the research agents that investigate methods and capabilities within the engine.

Tools

Tools are programs that perform specified operations: retrieving records, preparing texts, locating passages, or constructing measures.

Tools provide operations. Workflows arrange those operations, agents use their outputs, and reusable implementations can be retained as resources.

Resources

Resources provide the materials on which researchers and agents draw: source texts, accumulated findings, methodological guidance, and packaged toolkit components retained for revision and reuse.

These components include agent skills, which combine instructions, reference materials, and supporting tools for particular research tasks.

Evidence, findings, and methodological decisions are retained as resources for subsequent inquiry. This lets a study develop while preserving the reasons behind its methods and conclusions.

Development agents

Human researchers direct development agents in building and maintaining the deep-research engine. Research agents work within that engine to investigate methods and capabilities.

Deep-research engine

The deep-research engine coordinates continuing inquiry into capabilities, methods, and tools. It connects research needs and technological developments to decisions about what TextData should incorporate, modify, or build.

This is the infrastructure design. The shareable engine’s integration and evaluation remain proposed.

Research agents

Research agents examine literature, primary sources, and state-of-the-art technical guides or repositories. Questions from different perspectives guide retrieval and synthesis.

Under researcher direction, agents use this source-linked record to assess existing components and decide what should be adopted, adapted, or developed. Reported capabilities, proposed uses, and limitations need to remain distinct.

Tools within the engine

A suitable component may be incorporated directly; an existing tool may be adapted; or a method may need an implementation developed and tested. This draws on approaches that turn published methods into reusable tools.

Figure 1. TextData infrastructure design. Human researchers and development agents build the deep-research engine; AI research agents work within it to build tools, workflows, and resources to support research tasks. The engine is the adaptive core. Agent labels denote roles.

Process

The inquiry begins with the researcher and returns to the researcher’s judgment. Follow the path from a question to evidence retained for further work.

Define the inquiry

Human researcher

The researcher begins by defining the question and scope. These objectives remain the basis for configuring the analytical work.

The researcher sets the goal; the coordinator receives it.

Translate the goal into assignments

Coordinator agent

The coordinator translates the research goal into assignments for task agents, drawing on goal-directed research systems (Gottweis et al. 2026).

Workflows specify the sequence, conditions for proceeding, and how results pass between agents.

Consult, act and use the results

Task agents, workflows, resources and tools

Within workflows, task agents consult resources, select and call tools, and use the outputs to guide their next actions. This follows the reasoning–action approach of ReAct (Yao et al. 2023) and programmable agent interactions (Wu et al. 2024).

Tool outputs guide further action. Results return to the coordinator; a tool output alone is not a scholarly finding.

Bring the results together

Coordinator agent

The coordinator connects the contributions and returns integrated findings for the researcher to examine.

Integration makes the evidence available for assessment; authority over conclusions stays with the researcher.

Assess evidence and direct further work

Human researcher

The researcher assesses the evidence, judges which claims it supports, and directs any further work. The question, analysis, or instruments can be revised as the inquiry develops.

If further work is needed, it returns to the inquiry under researcher direction.

Retain what the inquiry has learned

Resources for subsequent inquiry

Evidence, findings, and methodological decisions are retained as resources for subsequent inquiry. This lets a study develop while preserving the reasons behind its methods and conclusions.

The next inquiry can draw on, reconsider, and revise these resources.

Adaptive core

The deep-research engine coordinates continuing inquiry into capabilities, methods, and tools. It connects advances in LLMs and open-source components with what the research needs.

Start with a need or a development

Research need or technological change

A research question may require a capability the current instruments lack. A model advance may warrant revising an existing procedure. A tool developed elsewhere may provide a capability that a project needs.

The engine connects investigation of these developments to decisions about what TextData should incorporate, modify, or build.

Investigate methods and capabilities

Research agents

Research agents examine literature, primary sources, and state-of-the-art technical guides or repositories. Questions from different perspectives guide retrieval and synthesis (Shao et al. 2024).

Findings accumulate with their sources in a shared knowledge vault, where they remain available for retrieval and reconsideration.

Textual-data literature Agent-design records

Assess what fits the inquiry

Agents under researcher direction

Under researcher direction, agents use this source-linked record to assess existing components and decide what should be adopted, adapted, or developed. Reported capabilities, proposed uses, and limitations need to remain distinct.

A component’s availability is a reason to investigate its fit, not proof that it supports the intended research claim.

Adopt, adapt or build

Tools, workflows and resources

A suitable component may be incorporated directly; an existing tool may be adapted; or a method may need an implementation developed and tested. This draws on approaches that turn published methods into reusable tools (Miao et al. 2026).

Adopt: Incorporate a suitable existing component into the research workflow.
Adapt: Modify an available component to meet the inquiry’s requirements.
Build and test: Develop an implementation when the required method or capability is not available in a suitable form.

The choice is governed by the research need and the assessment of the available evidence.

Examine the revision

Evaluation, including proposed model comparisons

For model updates, the proposed comparison uses the same passages with the new model and its predecessor before deciding whether to revise a workflow. Repeated trials of model-dependent tasks and checks after model or workflow updates are also planned.

Repeated-run and cross-version evaluations remain proposed. The existing component demonstrations do not establish end-to-end research performance.

Revise and retain

Continuing inquiry and development

The engine’s continuing inquiry informs the tools, workflows, and resources used in research. Development agents maintain and revise the engine that supports this work. Sources, findings, and methodological decisions remain available for later reconsideration.

New research needs and technical developments begin the cycle again. Adaptation is continuing research and development under researcher direction.

Deep-research engine

Different roles, shared researcher direction

Task agents carry out the study’s assignments. Research agents investigate methods and capabilities within the engine. Development agents build and maintain the engine. These labels describe roles, not separate levels of authority.

Reliability

The design addresses hallucination—unsupported or fabricated claims (Kalai et al. 2025)—inconsistency across runs (Yang 2026), and sycophancy: too readily endorsing a user’s position (Cheng et al. 2026).

Source links, software quotation checks, and separation of production from review follow verification-based approaches (Dhuliawala et al. 2024). Qualitative workflows seek alternative interpretations and contrary evidence, adapting constant comparison to make assumptions and contradictions explicit (Boeije 2002).

Software can repeat specified deterministic operations, while judgments depend on model outputs. Variation can arise within the system serving a model even with fixed prompts and settings (He and Thinking Machines Lab 2025). Repeated trials and checks for losses in previously demonstrated capabilities after updates are planned (Grace et al. 2026).

What the checks show—and their limits. Open a challenge to see its design response and the component evidence.

Unsupported claims

Design response. Link claims to source passages and separate production from review.

Evidence and limits. A fresh source reading corrected a reviewer’s claim in the engine work. This is a documented correction, not an error-rate estimate.

Engine evidence

False confidence or premature completion

Design response. Retain unknowns and researcher-defined stopping criteria.

Evidence and limits. A citation review declined to certify saturation; measurement specifications retain unknown fields. These are recorded decisions, not a general solution to false confidence.

Literature review Measurement design

Faulty interpretation

Design response. Separate quotation checks from judgments of meaning.

Evidence and limits. Synthetic actor substitution and omission errors pass the annotation source checks. A synthetic Gioia memo contains an unsupported negative-case inference. Exact quotation matching cannot detect these errors of interpretation.

Annotation checks Gioia example

Premature agreement or lost context

Design response. Seek alternative interpretations and contrary evidence, and preserve reasons for revision.

Evidence and limits. Interview interpretations changed during the Grounded Theory work. Review was unblinded and used the producers’ model; it does not establish independent validation.

Grounded Theory evidence

Changed scope or unauthorized action

Design response. Bind execution to the approved plan.

Evidence and limits. Maintenance software refuses a plan after its inputs change. This establishes a specific software check, not general control of agent behavior.

Maintenance checks

Inconsistency and model change

Design response. Preserve inputs; evaluate repeatability and version changes.

Evidence and limits. Annotation requests are retained. Repeated-run and cross-version evaluations remain proposed. Fixed prompts and model settings alone cannot ensure repeatable outputs.

Retained annotation inputs

Prompt injection, bias and leakage

Design response. No general safeguards demonstrated.

Evidence and limits. Withholding interview transcripts is case-specific. Broader safeguards against malicious instructions in source material, biased outputs, or disclosure of sensitive material require evaluation.

Interview-material boundary

Explore AI defects and the limits of the safeguards

Reuse and adoption

Researchers can begin with individual components, examine their sources and limitations, and adapt them to an inquiry. Sharing and maintenance through versioned open-source releases are proposed.

Find and assess a component

Searchable databases and source profiles organize methodological and technical research. Records distinguish reported capabilities from proposed uses and preserve sources and limitations so researchers can assess a component’s fit.

Literature database Agent-design databases Source profiles

Inspect and adapt the procedure

Several toolkits are packaged as agent skills: methodological guidance, agent instructions, templates, and supporting software. Agents can consult these procedures; researchers can inspect, adapt, and reuse them.

Grounded Theory Gioia Methodology Eisenhardt Case Methodology

Reproduce specified checks

Saved inputs and worked examples let users reproduce specified checks and examine how outputs were obtained. Reproducing a software check does not validate the interpretation or guarantee repeatable model judgments.

Acquisition comparison Measurement specifications Saved annotation responses

Contribute through future releases

Versioned open-source releases and community contributions are proposed. Researchers could adopt individual components, adapt them within their own workflows, and contribute source records, tools, tests, and corrections. Implementation files are being prepared for release.

Engine development