Research

Research Summary

This initiative studies the systems that assign, track, and recover agent work. The unit of work includes an objective, dependencies, an execution environment, acceptance criteria, and a handoff to the next stage.

OpenSymphony connects an issue source to coding runs and workspace management. GSD-2 examines external execution control. The task-graph article investigates structural checks against skipped steps, while dynamic workflows move coordination into executable code.

Getting Out of the Loop adds the operational questions that appear when a fleet runs with less direct supervision. We examine dependency gates, review loops, conflicting changes, stalled runs, retries, and escalation. Measures include accepted throughput, recovery time, duplicate work, review effort, and cost across the complete workflow.

Research Summary

This initiative studies Claude Code, Codex, OpenCode, OpenHands, Pi, Cursor, Windsurf, and other coding harnesses working in real repositories. We examine how they understand architecture, select tools, edit code, run tests, and respond to review. OpenSymphony connects these runs to issue-scoped workspaces and review gates; dependency management and fleet operation are covered by Work Orchestration.

Our Astra reasoning-effort experiment compares token use, runtime, and tool activity across matched tasks. The Claude model comparison examines whole-file writes and diff churn. Broader recommendations require more tasks and repeated trials.

spotlight-md provides persistent comment threads for human-agent editing. Our Codex provider-switch investigation shows how session metadata affects access to prior work. Together with Frontier Code Intelligence, these articles extend the scope to repository knowledge, feedback, and continuity.

We investigate which combination of model, context, instructions, and review produces an accepted change at a reasonable cost. Measures include correctness, architectural fit, test behavior, unnecessary edits, recovery, and the human effort needed to finish the work.

Research Summary

This research studies the infrastructure needed to keep agent systems usable over time. Scope includes local and remote deployment, runtime identity, tenant isolation, configuration drift, memory integrity, monitoring, and recovery.

The multi-gateway OpenClaw article examines credential boundaries. Managing OpenClaw with Claude Code documents repeatable operational skills. Memory repair and a crash postmortem provide concrete failure investigations.

Remote deployment extends the work to hosted environments. Tokenomics Part 1 and Part 2 investigate usage accounting and practical cost controls.

We examine how to reproduce an environment, detect drift, restore service after failure, and reconcile task-level usage with the bill. Runtime design for personal agents and inter-agent federation are covered by the neighboring initiatives.

Research Summary

Personal agents combine a user’s context, tools, memory, and unfinished work across conversations. This research compares how OpenClaw, Hermes, and other persistent runtimes support that relationship. Browser and desktop agents, including Operator and Manus, provide comparative examples of interaction with applications.

The Hermes and OpenClaw comparison examines gateway routing and learning-loop design. OpenAI dot explores what changes when an assistant retains context and continues work between conversations. Agent email reporting illustrates a concrete tool that lets an agent report progress while its user is away.

Research questions include how memory is selected and corrected, how an agent recognizes unfinished work, and how users inspect or stop background actions. We also examine portability of skills and context across local and hosted runtimes. Deployment, configuration, recovery, and billing are covered by Agent Runtime Operations.

Research Summary

This research compares tool integration, agent communication, and gateway federation. OGP, A2A, and MCP: Three Lanes, Same Highway defines their respective boundaries. CLI Tools vs MCP examines when a protocol server earns its integration cost.

The OGP Project Layer case study explores shared work across independent agents. Runtime-independent federation examines adapters beyond OpenClaw. The 100-build postmortem records identity and persistence failures encountered during implementation.

We investigate discovery, authenticated messaging, delegated authority, project state, and controlled relay into local runtimes. Tests distinguish a documented capability from a complete collaboration that survives restart, denial, retries, and peer loss. Governance of delegated actions is covered by AI Governance & Auditability; deployment and observability belong to Agent Runtime Operations.

Research Summary

This initiative connects operational governance of agents with oversight of the organizations that develop and deploy AI. Research questions concern authorization, retained evidence, accountability, conflicts of interest, and the ability to enforce a decision.

Federation Without Governance examines delegated authority across gateways. Five Layers of No describes checks on inbound OGP messages. These provide implementation cases for distinguishing a requested action from an authorized action.

Our frontier oversight analysis considers the evidence and authority available to independent reviewers. A Charter Is Not a Control asks which mechanism still operates after the subject stops cooperating. These articles supply governance arguments and open design questions.

We investigate records that support incident reconstruction, review scopes that expose material failures, and controls that bind delegated actions. Technical credential and execution protections are covered by Agent Security & AI Cybersecurity.

Research Summary

This initiative examines both the security of agent systems and the use of agents in cybersecurity work. Execution isolation, tenant boundaries, credential access, and network permissions determine what an agent can reach. Audit workflows determine whether suspected vulnerabilities and proposed repairs can be verified.

Agent Vault explores credential mediation. Our Grok Build investigation examines repository exposure and tool supply-chain risk. The OpenAI-Hugging Face incident analysis connects agent behavior to credential controls, safe stopping, and tamper-evident records.

The AI Cybersecurity Playbook is the practical resource for this work. Its implementation article covers path-aware audits, penetration testing, fix verification, and incident response.

We investigate how to give agents enough access for a defined task while containing unrelated resources, how to preserve investigation evidence, and how to verify that a repair closes the reported attack path. Governance of authority and independent oversight is covered by AI Governance & Auditability.

Research Summary

This initiative studies the path from source documents and fragmented data to knowledge that teams can query, maintain, and use in decisions. The publication series spans legacy OCR, contract intelligence, structured data aggregation, GraphRAG, iterative discovery, and Entropy’s operational knowledge architecture.

Our contract intelligence work examines extraction that retains document hierarchy and supports questions about clauses and obligations. One Document, Three Truths extends that work to isolated workspaces with team-specific extraction schemas for Legal, Procurement, and HR.

GraphRAG and Multi-Agent Deep Research Architecture investigate relationships and accumulated knowledge as inputs to retrieval and further research. The Entropy architecture connects a maintained knowledge base to customer strategy, simulation, and action.

Research questions include:

• How accurately do extraction pipelines preserve tables, document hierarchy, obligations, and relationships across poor-quality source material?
• When do graph retrieval, structured queries, or document retrieval provide the most useful evidence for an answer?
• How should systems track provenance, resolve conflicting information, and refresh knowledge as sources change?
• How can shared infrastructure support different extraction needs while enforcing workspace and tenant boundaries?
• How do teams evaluate whether generated answers and playbooks remain grounded in the underlying records?

These questions connect document processing to operational knowledge: a system must preserve useful evidence, make it accessible to the right team, and keep it current enough to support decisions.

Research Summary

This initiative examines how AI systems turn business context into useful actions inside existing team workflows. Our publications cover account research and sales risk, customer support, employee coaching, meeting assistance, social content, and automated change and architecture reviews.

The sales risk pipeline case study describes a working system that enriches Notion account records with web evidence, exposes recommendations through an API, and lets a conversational agent draft outreach and prepare presentations. It provides an implementation example; sustained business outcomes require measurement over time.

The meeting copilot proposal explores routing meeting context to specialist agents. Agentic Automation for Social Content examines creation, approval, and scheduling with n8n and Airtable. Together with our support and coaching work, these publications inform several research questions:

• How should an agent obtain the context and tools needed for a particular business task?
• Which actions can run automatically, and where should a person review a recommendation or approve execution?
• How do teams measure time saved, recommendation accuracy, failed actions, and operating cost?
• Which integrations and workflow patterns can another business unit reuse, and what must remain specific to its work?

The research follows workflows from a working prototype through integration, evaluation, and transfer to the team responsible for operating them.

Research Summary

This initiative asks when training or adaptation improves a deployed system enough to justify its data, compute, and maintenance cost. Options include fine-tuning, adapters, retrieval, prompt optimization, model routing, and reinforcement learning around agent behavior.

The Limits of Fine-Tuning records a domain-specific capability regression and the resulting hybrid inference architecture. Our agent RL overview and framework comparison examine optimization loops around tools and workflows. The Nova Forge assessment considers training claims and operational economics.

We investigate data quality, reward design, held-out task performance, and regression on general capabilities. Comparisons include retrieval and routing baselines, training and serving cost, and the effort needed to reproduce the adapted system. Open Model Strategy & Economics covers selecting and deploying base models.

Research Summary

This research compares open-weight models and hosted frontier services for coding, reasoning, retrieval, and multimodal work. Qwen, Kimi, DeepSeek, Gemma, and other model families are candidates for particular workloads. Evaluation includes the model’s serving route, context limits, availability, hardware requirements, and data access constraints.

The Qwen and Kimi repository-analysis comparison combines a shared task with blind review. The Gemini pricing analysis examines how token efficiency changes effective task cost. Tokenomics Part 3 considers open-model economics and the difference between API and subscription pricing.

We ask which model and deployment route meet a task’s quality threshold at the lowest measured total cost. Repeat trials, accepted outcomes, fallback behavior, cold loads, and sustained latency help answer that question. Model adaptation and capability retention are covered by Training & Adaptation.

Research Summary

This initiative studies how to evaluate model outputs and bounded decisions. G-Eval remains a method for criteria-driven assessment without a reference answer. Our work also includes retrieval benchmarking, human review, and a code-review prompting experiment.

Jev’s Open Rivals extends the scope to typed decisions, probability calibration, and selective automation. It compares computational routes and examines published judge studies. Confidence thresholds need held-out evidence of coverage and accepted-case error; a fallback can repeat the first model’s errors.

Research questions include how to construct representative datasets, identify judge bias and disagreement, and separate model fitting, calibration, threshold selection, and final testing. We also measure the full decision route, including fallback cost and latency. A changed question schema, model, or acceptance threshold changes the policy being evaluated.

Research Summary

This initiative evaluates how agents execute real workflows. We examine individual task performance and system-level coordination through completion evidence, tool errors, recovery, handoffs, and the cost of reaching a verified result.

A Practical Guide to LLM & Agent Evaluation provides the evaluation framing. Taming Tool Calling with Kimi K2.5 investigates tool-surface reduction, structured guidance, and hybrid routing. The MiniMax experiment inside OpenSymphony connects review, tests, latency, and recovery to an actual engineering run.

Engineering Behavior Observatory is the research platform for preserving native trajectory evidence and workspace outcomes, applying behavioral assessments, and comparing repeated trials in the Behavior Atlas. Human review and capture-quality checks help establish which conclusions the recorded evidence supports.

We investigate how different harnesses, models, and workflow designs affect error recovery, objective retention, verification, and unnecessary work. Comparisons hold the task and acceptance criteria steady and report variation across runs.

Research Summary

This initiative examines the full path from source material to a rendered media asset. It includes narrated explainers, image generation, text-to-video, music models, and reusable production tooling.

Our article-to-video implementation replaces rigid template filling with compositions authored for each article. The narration timing postmortem traces clipped speech to a mismatch between requested duration and generated audio. Hyperframes explores how encoded production structure supports agent-driven rendering.

The MiniMax H3 study examines reference-controlled scenes and continuity limits. Engineering Determinism for Image Generation investigates checks around generated visual outputs.

We evaluate prompt adherence, readable text, scene coherence, audio alignment, render failures, and the cost of a usable result. Model specifications help select candidates; produced assets and repeatable checks establish whether a workflow meets its requirements.

Research Summary

This initiative studies how models interpret and retrieve information across text, images, and audio, and how those capabilities support interactive systems. Realtime voice and video remain one strand alongside visual reasoning, multimodal search, and long-context processing.

Late Interaction: ColBERT to Wholembed v3 examines retrieval representations and fine-grained matching. The Qwen 3.5 deep dive explores native multimodality in smaller open models. Our voice-avatar work connects a chatbot’s knowledge to a visual, voice-enabled interface.

Research questions include whether visual or audio context improves answers on a defined task, how retrieval changes relevance, and how latency, context size, and device constraints affect interaction. Live and ambient systems also need explicit decisions about what context to capture, retain, and expose to tools. Image, video, and music production are covered by AI Media Production Workflows.

Research Summary

This completed study compares two ways to produce educational visuals: generating code for diagrams, animations, and simulations, and generating bitmap illustrations and assets. The published guide describes tools and production approaches.

Code-driven tools are useful when positions, labels, equations, or timing need explicit control. Generated illustrations can supply subjects, textures, and other assets around that structure. The choice depends on the concept and the precision it requires.

Review covers factual correctness, label readability, consistency, and the relationship between a visual and its learning objective. The study documents production methods; a claim about improved retention would require a separate learner evaluation. Broader media tooling is covered by AI Media Production Workflows.

Research Summary

This initiative investigates personalized feedback, assessment generation, and learning paths. The AI tutor overview establishes the pedagogical framing; implementation articles examine cost and validation.

The cost-aware tutor case study studies a constrained response format and a simpler workflow. DSPy Unleashed explores a system that generates and improves educational questions. LearnLens addresses discovery and analysis of video learning material.

We investigate whether feedback is accurate, whether generated assessments match the intended skill, and whether a learning path helps a learner make progress. Measurements of answer quality and inference cost inform engineering decisions. Claims about learning gains require learner studies with appropriate comparison groups and retention measures.