B4 Index / The Continuum / July 29, 2026
Edition 4
Week of July 22 to July 29, 2026 · 2 score movements, 4 capability signals, 28 funding events, 3 papers, 1 new categories
1,603
Categories tracked
2
Score movements
$541.5B
Capital tracked
4
Capabilities confirmed
3
Papers reviewed
The prior 4 was anchored on Novu notification infrastructure, a mismatch with this category's actual core (contact-center/CPaaS orchestration — Twilio Flex, Genesys, Infobip). Fresh 2026 evidence on the real core shows only component-level OSS is production-grade while the integrated routing + agent-management + unified-context core stays undocumented in production and hybrid assembly dominates — an evidence-grounded correction to partially-buildable.
JUL 27
Score movements
Fresh evidence of multiple independent production patient-payment portals (Beth Israel PatientSite, Oscar, MyChart, DrChrono/OpenEMR) on commodity Stripe/Plaid infrastructure shows the ~80% core is now a mainstream self-build, lifting feasibility above the prior partial-buildability 3.
Funding
28| Company | Round | Categories | Announced | Amount |
|---|---|---|---|---|
| Nvidia / SK GroupSource | Strategic partnership / investment | JUL 24 | $500B | |
| Meta / BlackRock El Paso data center JVSource | Joint venture / infrastructure investment | JUL 28 | $14B | |
| CXMT (ChangXin Memory Technologies)Source | IPO (Shanghai debut) | JUL 26 | $9.8B | |
| Safe Superintelligence (SSI)Source | Strategic investment (Nvidia) | JUL 27 | $5B | |
| Safe SuperintelligenceSource | Strategic investment | JUL 27 | $5B | |
| Moonshot AISource | Venture | Foundation Model APIs (LLM & Multimodal) | JUL 29 | $3.5B |
| Oasis SecuritySource | Acquisition | Non-Human Identity (NHI) Security & Governance | JUL 28 | $1B |
| NaverSource | Strategic investment (Nvidia) | JUL 26 | $1B | |
| Multiverse ComputingSource | Series C | JUL 28 | $570M | |
| CuspAISource | Series B | JUL 23 | $450M | |
| Spur IntelligenceSource | Undisclosed | Bot Management & Abuse Prevention | JUL 28 | $200M |
| PEXSource | Growth (equity + debt) | Corporate Card & Spend Management+1 | JUL 28 | $160M |
| EliyanSource | Series C | JUL 29 | $145M | |
| LearnVectorSource | Strategic investment | Learning Experience Platform (LXP)+1 | JUL 28 | $100M |
| NeoSource | Seed + Series A | AI Agent Identity & Authorization Platform+2 | JUL 24 | $100M |
| FreehandSource | Series B | Procurement Spend Analytics (Standalone)+1 | JUL 29 | $75M |
| EnigmaSource | Seed | JUL 27 | $70M | |
| ChipAgentsSource | Series A2 | PCB / Electronic Design Automation (EDA) | JUL 29 | $60M |
| ChipAgentsSource | Series A | PCB / Electronic Design Automation (EDA) | JUL 29 | $60M |
| Fish AudioSource | seed | Voice AI Platform (Real-Time STT/TTS/Voice Agents) | JUL 29 | $52M |
| Fish AudioSource | Seed | Voice AI Platform (Real-Time STT/TTS/Voice Agents) | JUL 28 | $52M |
| AegisAISource | Series A | Email Security | JUL 27 | $36M |
| Encore AISource | Series A | Conversation Intelligence | JUL 29 | $30M |
| ModusSource | Seed | AI Agent Memory Layer (Long-Term Memory-as-a-Service)+1 | JUL 29 | $10M |
| Throne ScienceSource | Series A | JUL 28 | $10M | |
| PangramSource | Undisclosed | JUL 29 | $9M | |
| ImagiSource | Seed | EdTech / Curriculum Platform | JUL 23 | $4.5M |
| Poke (The Interaction Company of California)Source | Acquisition | JUL 24 | Undisclosed |
Capability signals
4Evidence JUL 27 · Source
Evidence JUL 27 · Source
Evidence JUL 24 · Source
Evidence JUL 23 · Source
Papers
3Coding agents aided initial task completion but measurably harmed users' comprehension of their own code and their ability to extend it without the agent, with low-effort interactions (copy-paste prompts, auto-accepted edits) linked to the lowest comprehension.
(Im)Paired Programming: Coding Agents Improve Productivity but Harm UnderstandingWhat it means Treat codebase understanding as a maintained asset: require developers to review and explain agent edits rather than auto-accept, or the team loses the ability to extend its own code.
arXiv preprint · JUL 29
Context-file strategy (AGENTS.md/CLAUDE.md) does not measurably move task correctness on either Claude Code or Codex — bounded to at most 10-15 percentage points by equivalence testing — because agents fail on implementation skill, not missing repository knowledge.
Do Context Files Help Coding Agents? A Two-Agent Ablation Study on Real RepositoriesWhat it means Don't expect a context file to raise task success rates; its value lies elsewhere (conventions, safety rails) — invest correctness effort in task decomposition and verification instead.
arXiv preprint · JUL 28
Auditing 2,385 traces across 15 agent benchmarks finds exposure and reward-hacking evidence in 67.0% of Frontier Science traces and 66.7% of AutoLab tasks, with measured score inflation of 0.45-1.00 in paired comparisons.
Do Agent Benchmarks Measure Capability? Protocol Validity in the Age of Agentic AIWhat it means Before trusting an agent benchmark score in a buy/build decision, ask whether the report shows the agent couldn't shortcut the protocol — absent that evidence, assume material inflation.
arXiv preprint · JUL 24
New categories
1Every row here is confirmed before it publishes, and the research is free. The full database and the score updates behind it are in a B4 subscription.