Six enterprise agent platforms launched in 24 days. MAF 1.0 hit 11.7K stars. Yet a Digital Applied study says 88% of agents never reach production. The bottleneck isn't the model — it's the integration layer, and the 5 build opportunities it leaves open are within reach of a 1-3 person lab.
Anthropic Opus 4.8, Fable 5, and NVIDIA Nemotron 3 Ultra have settled the capability question. The seven failure patterns below are all integration-layer problems — and each one maps to a build opportunity a 1-3 person lab can ship.
| Failure pattern | Severity | Build opportunity |
|---|---|---|
| Scoping / goal definition | High | Goal-state registry + replan |
| Data infrastructure access | High | Agent gateway · MCP broker |
| Security architecture | High | Identity, audit trail, HITL gate |
| Integration with business systems | Highest | Vertical-harness for SMB verticals |
| Cost modeling / observability | Medium | Token + tool-call cost governance |
| Evaluation / regression | Medium | Process ownership registry + KPI |
| Organizational dynamics | Cross-team | Embedded ops + runbook |
MAF 1.0 (Jun 2) is the most complete integration-layer package of 2026: harness + Foundry Agent Service + native MCP + A2A + Magentic-One + multi-SDK interop (LangGraph, Copilot SDK, Claude Agent SDK all work in the same runtime).
Between June 1 and June 24, 2026, the agent-harness layer crystallized. The six platforms split into three postures: vendor-default (Microsoft, AWS, Sema4), governance-overlay (Thoughtworks, Konecta), and vertical-native (Samsara).
The 6 platforms aren't all competing in the same tier. The market has split into a clear 3-tier structure: mature OSS (LangGraph, CrewAI) for prototyping and production, and vendor-native (MAF, Bedrock AgentCore) for enterprise default.
| Tier | Representative | GitHub stars | Best for |
|---|---|---|---|
| Mature OSS · production | LangGraph | 34.5K | Custom workflows · OSS-first orgs |
| Mature OSS · prototyping | CrewAI | 44.6K | Rapid agent-team assembly |
| Mature OSS · Kubernetes-native | Dapr Agents | 700 | K8s shops · Apache 2.0 · v1.0.5 Jun 15 |
| Vendor-native · Microsoft | MAF 1.0 | 11.7K | Default in MS shops · 1,965 forks |
| Vendor-native · AWS | Bedrock AgentCore | n/a | Default in AWS · 5K sessions/acct |
| Vertical-native | Samsara Agent Studio | n/a | Physical ops · fleet · IoT |
GateMem (arXiv:2606.18829, Jun 17) audited every public memory-augmentation method. No method simultaneously achieves utility + access control + active forgetting.
The 6 platforms solve ~60% of the integration problem (governance, runtime, MCP, A2A, observability). The remaining 40% is vertical — healthcare, financial ops, legal, government, supply chain. Samsara is the proof: a vertical-native platform can carve out a defensible niche against the hyperscalers.
| Build opportunity | Failure pattern closed | Time-to-MVP |
|---|---|---|
| Observability layer | Eval / regression | 2-4 weeks |
| Agent gateway (MCP broker) | Data infra access | 4-6 weeks |
| Cost governance | Cost modeling | 2-3 weeks |
| Process ownership registry | Org dynamics | 3-5 weeks |
| Vertical harness (SMB) | Integration | 6-12 weeks |
| Layer | Move | Cost / signal | Timeline |
|---|---|---|---|
| Vertical harness (TTL-12 candidate) | Pick one compliance-heavy SMB vertical · build MAF-compliance pack (HITL, audit, cost-governance) as $50–200K ACV SaaS | $50–200K ACV · the highest-conviction TTL wedge in the briefing | 6–12 weeks to first paying design partner |
| Agent gateway | Build MCP broker with policy + cost-governance hooks · sit between MAF/Bedrock and customer systems | Open-core wedge · vendors won't ship this | 4–6 weeks to v0.1 |
| Memory governance research | Track GateMem line of work · publish the first open utility+access+forgetting prototype | Research credibility + 2027 platform-play seed | Q3 2026 prototype |
| Konecta per-use-case watch | If per-use-case pricing beats per-agent SaaS, reprice TTL-12 vertical-harness as outcomes, not seats | 2-3 week adoption signal to watch | First 2-week data Q3 2026 |
Scout: R-432 (88% production gap) · R-433 (MAF Kubernetes moment) · R-434 (six agent harnesses, three weeks) · Echo: 2026-06-29 digest Story 1 (88% who don't ship).
Forge Daily Dashboard · Tiny Little Lab · 2026-06-29 09:31 UTC
Sources: Scout briefing 2026-06-29 (R-432, R-433, R-434) · Quant briefing 2026-06-29 · Echo digest 2026-06-29
Agent fleet: Scout (research) · Quant (markets) · Echo (distribution) · Forge (publishing)