AI research systems and agent design
Research foundation
A survey of AI research systems and reusable infrastructure, with a focused database of current agent-design and orchestration approaches.
What has been built
Two databases, with browsing and query tools, document systems and agent designs. A worked record connects a proposed restart test to reported repeated tool calls; it is not a performance test.
What comes next
Use the records to develop and evaluate TextData configurations; update them as agent methods change.
From research task to design choice
Use them to weigh technical options for a research task: what a component does, what supports it, where it stops, and what would have to be tested before TextData relies on it.
22,667
typed entries in the systems survey
A survey of research items and evidence units across 14 functions; not a count of unique systems.
34
entries in the agent-design database
Seven categories of design question; 19 entries rest on selected primary text and 15 on documented source summaries.
Browse the design assessments, compare mechanisms and inspect their sources. Use the survey to find related evidence across research functions.
Example: can a task restart safely?
AD-014 connects a proposed restart test to an unconfirmed report of repeated tool calls. Explore the records below; the test itself has not been run.
In the agent-design database
AD-014 · LangGraph persistence
- What it does
- Separates thread-state checkpoints from cross-thread application data stored in a store.
- Limits
- Restart durability depends on the backing store and configuration. Checkpoints do not make arbitrary external effects atomic or establish analysis validity.
- Possible use in TextData
- Keep execution position separate from the enduring study evidence and decisions.
- Question to evaluate
- Interrupt a task and check both correct resumption and absence of duplicated external effects.
In the systems survey
7 survey entries share source addresses with AD-014. A shared address is a connection to inspect, not independent corroboration.
The reporter described a checkpoint sweep starting a second copy of a long-running call while the first was still active.
This production report remains unconfirmed. It identifies a possible failure and gives no estimate of how often it occurs.
Inspect the other 6 source connections
The record states the mechanism, its limit, a possible use and the test it would need; the survey entries supply selected project paraphrases and source addresses for inspecting that judgment. Neither shows that the component works in TextData.
Open the agent-design appendix
The agent-design database
Papers, specifications, engineering accounts and reusable components, grouped by the design question they bear on. Each entry keeps what the component does, what its evidence shows and where it stops apart from a possible use in TextData and the question an evaluation of that use would have to answer.
Categories
Choose an entry to read it in full. Compare sets the entries shown side by side.
Action and planning
How should an agent choose, execute and revise its next step?
Organizes language agents through memory, available actions and decision cycles.
Interleaves model reasoning, tool actions and observations from the environment.
Stores verbal feedback from prior attempts and uses it to guide a later attempt without updating model weights.
Searches alternative intermediate states using model-based evaluation and tree search.
Plans tool calls with dependencies, dispatches independent calls in parallel and can replan after observations.
Orchestration and reuse
Which responsibilities belong in program structure, and which in a model decision?
Distinguishes predefined execution paths from model-directed steps; describes routing, parallelization and orchestrator–worker patterns.
Separates declared input/output interfaces and modular programs from prompts optimized against a metric.
Serializes agent and workflow definitions and connects them to runtime adapters.
Runs an agent loop inside a durable workflow and records model and I/O operations as replayable activities.
State and context
What must persist, and what does the model need to see now?
Uses references, selective retrieval and compaction to manage what a model sees during long tasks.
Separates the durable session event log, the agent execution loop and the sandbox.
Separates thread-state checkpoints from cross-thread application data stored in a store.
Moves information between limited working context and external recall/archival storage through explicit function calls.
Extracts and retrieves persistent information; the inspected paper distinguishes extraction from update decisions.
Tools and skills
How should capabilities and methodological instructions be exposed?
Resources expose contextual data; tools expose callable operations through a shared protocol.
Packages instructions with optional scripts, references and assets, with progressive loading of detail.
Designs the agent–computer interface around search, a bounded document view, editing and feedback.
Delegation and handoff
When does dividing work help, and what must pass between workers?
Configurable agents exchange messages under programmable conversation and termination rules.
Converts a scientist’s goal and constraints into a plan with specialized generation, critique and refinement work.
A lead agent delegates bounded searches, receives findings and decides whether to continue.
Classifies system-design, inter-agent alignment and task-verification failures.
Compares repository-only, raw-trace, summary and structured-note views at frozen worker-takeover points.
Evaluation and adaptation
How will we assess a change without confusing extra computation with better design?
Develops evaluators through user inspection and revision of criteria.
Connects evaluation tasks, solvers, scorers and inspectable per-run/per-sample logs.
Assesses function calling, including state and viable call sequences in multi-turn tasks.
Evaluates repeated success in stateful, policy-bearing interactions among a user simulator, agent and tools.
Uses execution feedback to reflect on and evolve prompts while selecting among candidate trade-offs.
Optimizes module-level instructions and demonstrations against downstream task scores.
Frames agent assessment jointly in terms of task accuracy, cost and intended generality.
Evidence and authority
How can we recover the basis of a claim and keep source content within its proper role?
Supports concept induction with analyst inspection, editing and re-analysis.
Represents entities, activities, derivation and responsibility.
Defines shared names and relationships for records of model and agent operations.
Combines legitimate tool tasks with attacks through untrusted content and evaluates utility and attack outcomes.
Examines privacy-relevant behavior in agent actions as well as responses about privacy norms.
No entry matches. Clear the search or choose another category.
The entries could not be loaded.
The systems survey
Find research items by function, setting and evidence state. Each entry records unresolved questions; a gap does not show that a capability is missing.
Search names, IDs and available notes, open a record, or download the matching index rows.
19,401 summary passages are omitted for privacy or reuse reasons. The underlying evidence state and gaps remain visible.
Research functions
Filters
The matching index:
Where the evidence stops
- The databases, browsing tools and worked reading are complete. No listed system was installed or benchmarked for TextData, and no TextData configuration has been evaluated.
- The next step is to build and evaluate configurations using these records, including whether evidence survives delegation and contrary passages prompt revision.
- Coverage is bounded and the design collection is selective. Entries are not unique systems; function counts overlap, and repeated sources are not independent findings.
- Evidence states describe documentary assessments, not a new independent check of every primary source. Proposed uses and evaluation questions remain proposals.
- 27 of the survey’s 22,667 entries are not shown on this page.
Credits
- Records credit the authors and maintainers through their public source addresses. Third-party publications, software and documentation retain their own terms.
- Project code: Apache 2.0. Project documentation and assessments: CC BY 4.0. Source notices and reusable files are in the component folder.