Research Roadmap

Last updated: 2026-07-28

Open questions and areas requiring deeper investigation. These are the unknowns that, if answered, would unlock higher-confidence decisions or stronger content.

---

High Priority

1. Cost Economics of Background Agents

Question: What does it actually cost to run a Devin/Cursor/Claude agent workflow at various scales (solo dev → small team → startup)?

Why it matters: The background agent pattern currently lacks hard numbers. Real cost data would make the argument concrete and help engineers decide when to use agents vs. traditional approaches.

**Sources:

Devin pricing (current)

Cursor agent usage data

Claude API costs per task

Time saved vs. cost incurred (from the Devin case study)

Status: Not started

---

2. Testing Ownership in Agent-Produced Code

Question: If an agent writes the code, who writes the tests? What testing patterns work best for agent-produced PRs?

Why it matters: The current cluster covers code generation and trust verification but doesn't address testing quality. This is the biggest blind spot.

**Sources:

The Devin case study (did Devin write tests?)

Cursor agent testing behavior

Human review patterns for agent tests

Status: Not started

---

3. Team Dynamics Under Async Agent Workflows

Question: How do team communication patterns change when some engineers work asynchronously with agents? Does it create coordination overhead? Does it scale?

Why it matters: The current content is individual-focused. To reach wider audience, understanding team/org implications is critical.

**Sources:

Case studies of teams using Devin/Cursor agents

Pair programming vs. agent pairing comparisons

Status: Not started

---

Medium Priority

4. Graduated Trust Feasibility on Hypermedia

Question: Can the graduated trust web-of-trust system be implemented using Seed Hypermedia capabilities and profile attestations? What would the schema look like?

Why it matters: Validating that the proposal can be built on HM would give it a concrete home and differentiate it from theoretical trust frameworks.

Status: Not started

---

5. State Machine Formalization for Common Agent Patterns

Question: What are the canonical state machines for common agent workflows (research→code→test→PR, bug→fix→verify, etc.)? Can we abstract them into a library?

Why it matters: A formal model makes the background agent pattern teachable and implementable.

Status: Not started

---

Completed / Resolved

(None yet — first analysis)

---

How to Use This

When new evidence arrives (new articles, case studies, data points), check which questions it answers and update the status. Remove or downgrade resolved questions; add new ones as they surface.