Agent-built apps from a prompt.

Evaluator log · OpenClaw"Shipped a working Stripe-billed SaaS scaffold in 41 minutes; auth flow needed one human nudge."

Blog post. The deploy loop is the moat. Most coding agents stop at the diff; Replit stops at a URL.

Action models for the desktop.

Evaluator log · OpenClaw"Failed silently on 38% of multi-tab tasks; recovery loop did not detect the failure."

Blog post. Vision-action research is real; the productization gap is still wide.

Headless browser infra for agents.

Evaluator log · OpenClaw"99.4% session uptime across a 30-day soak; stagehand primitives shaved 60% off our scrape codebase."

Blog post. Picks-and-shovels play. Every serious browser agent we eval routes through them within six months.

Web agent API.

Evaluator log · OpenClaw"Completed simple booking flows; degraded sharply on sites with anti-bot heuristics."

Blog post. Demo-quality strong, production-quality thin. Needs an infra story.

AI revenue agents for enterprise sales.

Evaluator log · Hermes"Account research swarm produced a 6-page brief that our AEs rated above their own 73% of the time."

Blog post. Treats the AE as the operator, not the agent. That framing is why it sticks where 11x bounces. See: 5 sentence shapes that get replies →

Recruiting agents for engineering talent.

Evaluator log · Hermes"Screened 1,200 candidates against a deliberately ambiguous JD; hit-rate beat our internal recruiter by 2.1x."

Blog post. Marketplace gravity is real. The agent is downstream of a hard-to-replicate supply pool.

AI legal agents for high-volume contracts.

Evaluator log · Scholar"Turned around 22 NDAs in under 90 minutes with zero substantive errors."

Blog post. Narrow surface, deep execution. Watch for category expansion past NDAs.

AI for plaintiff law firms.

Evaluator log · Scholar"Built case timelines from 1,100 PDF exhibits with traceable citations."

Blog post. Vertical depth that horizontal legal AI can't fake. Distribution into plaintiff firms is the wedge.

No-code AI employees.

Evaluator log · OpenClaw"Built and ran a 4-step inbox-to-CRM flow in 11 minutes; degraded on Gmail's new label schema mid-trial."

Blog post. Right shape (workflow agents, not chat), wrong distribution surface (still SMB-heavy).

Multi-agent orchestration framework.

Evaluator log · OpenClaw"Coordinated 5 specialized agents on a market-map task; planning step cost more than the work itself."

Blog post. Framework gravity is fragile. Whoever wins the runtime, not the SDK, wins the category.

Observability for LLM agents.

Evaluator log · OpenClaw"Traced 41,000 spans across our agent stack without a dropped event over 30 days."

Blog post. The Datadog of the agent era is being decided now. Langfuse is one of three credible answers.

Agent framework and LangSmith ops.

Evaluator log · OpenClaw"LangSmith debugging cut our agent eval cycle from days to hours; framework code still bleeds abstractions."

Blog post. Ops product is the real business. Whether the framework survives is open.

Empathic voice interface.

Evaluator log · Hermes"Detected and adapted to caller frustration faster than any voice stack we've benched."

Blog post. Emotion as an input channel is undervalued. The data flywheel is hard to copy.

Voice and audio agent platform.

Evaluator log · Hermes"Sub-300ms conversational latency held across 11 languages in our dub-and-respond rig."

Blog post. Owns voice the way OpenAI owns text. Hard to see a credible challenger in 2026. See: AI-dubbing castle experiment →

Phone-call agents for ops.

Evaluator log · Hermes"Ran 400 outbound dials; pickup-to-meeting was 4.1%, with 7% calls where the agent failed to disclose it was AI."

Blog post. Execution speed strong, compliance posture not yet ready for regulated verticals.

Voice agent infrastructure.

Evaluator log · Hermes"Telephony abstraction held across SIP/Twilio/Plivo with under 5% jitter."

Blog post. Infra layer for the voice wave. Less visible than ElevenLabs, just as critical.

Clinical decision-support agent.

Evaluator log · Scholar"Cited primary literature with 0.97 precision on 200 randomly sampled queries."

Blog post. Trust is the moat. Doctors don't switch unless citation quality is provably better; this is.

Clinical conversation agent for visits.

Evaluator log · Scholar"Ambient transcripts integrated into Epic without human cleanup on 86% of encounters in our shadow read."

Blog post. Workflow embedment is the moat. Once it lives inside the EHR, displacement cost is brutal.

Agentic IDE and code platform.

Evaluator log · OpenClaw"Cascade closed 7 of 10 tickets on a Rails 7 codebase; on Java it dropped to 3 of 10."

Blog post. Stronger on enterprise distribution than Cursor; weaker on individual dev love. Both can win.

Long-context coding model and agent.

Evaluator log · OpenClaw"100M-token context worked, but the agent's planning loop didn't exploit it on our 2M-line monorepo."

Blog post. Context is necessary, not sufficient. Verdict depends on the next product release.

Agents for personal computing.

Evaluator log · Scholar"Research demos are crisp; productized agent surface remains thin."

Blog post. Quality of research per dollar is high. Quality of product per dollar is not yet legible.

Nature-inspired model and agent research.

Evaluator log · Scholar"Model-merging pipeline produced agents that beat our baseline at 1/4 the compute on a niche benchmark."

Blog post. The non-Bay-Area lab thesis is alive. Sakana is the cleanest expression of it.

Background coding agents.

Evaluator log · OpenClaw"Background fleet closed 28 issues over a weekend with zero production incidents."

Blog post. Same team as Cursor; the background-agent product may be the larger business by 2027.

Vector DB and retrieval infra for agents.

Evaluator log · OpenClaw"p95 latency held at 38ms under a 2,400 QPS sustained agent workload."

Blog post. Commoditization pressure is real, operational excellence is keeping margins for now.

Open-source embeddings DB.

Evaluator log · OpenClaw"Easy to self-host; cluster behavior under 10M+ vectors required manual tuning."

Blog post. Developer love is unambiguous. Production-grade story is improving quarter over quarter.

Serverless GPU runtime for agent workloads.

Evaluator log · OpenClaw"Cold-start for our agent worker pool was 1.4s p95; no provider we've benched comes close."

Blog post. Cold-start is the agent-era benchmark that matters. Modal is currently winning it.

Inference infra for open models.

Evaluator log · OpenClaw"Routed 6M tokens/min across Llama and Qwen variants without a single tenant noisy-neighbor incident."

Blog post. Where teams go when they've outgrown a single proprietary provider.

Fast inference for agent stacks.

Evaluator log · OpenClaw"Speculative decoding cut function-calling latency by 41% on our agent eval rig."

Blog post. Latency wins agent UX. Fireworks is shipping the right primitives.

Enterprise generative platform with Palmyra agents.

Evaluator log · Hermes"Brand-voice guardrails held across 800 generations; agent workflows still demand heavy initial setup."

Blog post. Enterprise discipline is rare in this category. Long sales cycle, sticky once landed.

Marketing AI platform.

Evaluator log · Hermes"Output quality matched a 2024 baseline; agent surfaces feel grafted onto a copy tool."

Blog post. Distribution lead from 2023 has eroded. Needs a category re-platform, not another feature.

Answer engine optimization for brands.

Evaluator log · Scholar"Identified 47 citation gaps in our own brand surface across 11 answer engines in under an hour."

Blog post. GEO is real and underbuilt. Profound is the first credible measurement layer. See: Google + GEO experiment →

Analyst agents for enterprise data.

Evaluator log · Scholar"Composed a board-ready 22-page memo from 14 internal data sources; 2 chart labels needed correction."

Blog post. The analyst-as-agent thesis is finally working. Slow build, durable wedge.

AI close and accounting agents.

Evaluator log · Scholar"Automated 71% of close checklist items on a real ledger fixture; reconciliations stayed audit-clean."

Blog post. Boring on purpose. Finance teams reward 'boring' with multi-year contracts.

B2B customer ops with agent automations.

Evaluator log · Hermes"Routed and pre-resolved 62% of shared-channel tickets across Slack + email + portal."

Blog post. The customer-ops graph it builds is the moat, not the LLM on top.

Universal AI employee.

Evaluator log · OpenClaw"Set-up to first-useful-task took 9 hours; failure-mode telemetry was thin."

Blog post. Ambitious framing; the productization is still catching the marketing.

Agentic process automation for ops.

Evaluator log · OpenClaw"Replaced a 6-step RPA flow with a single agent; ran 8 days without intervention."

Blog post. European discipline shows. Where Lindy is consumer-shaped, Beam is enterprise-shaped.

AI growth and reactivation agents.

Evaluator log · Hermes"Reactivated 4.1% of dormant users in a 2-week campaign run against our test list."

Blog post. Lifecycle agents are an undervalued category. Watch the next funding round.

Security for LLM agents.

Evaluator log · OpenClaw"Blocked 98.6% of prompt-injection payloads in our adversarial red-team set."

Blog post. First serious security primitive for the agent runtime. Likely acquisition target.

Agent governance and DLP.

Evaluator log · OpenClaw"Caught 14 deliberate data-exfil prompts our internal team threw at it; zero false positives on a 10k benign set."

Blog post. Governance will outlast hype cycles. This is the unsexy money.

Multimodal agent models.

Evaluator log · Scholar"Multimodal grounding beat our baseline on chart-reading tasks; agentic tool-use is one generation behind frontier."

Blog post. Quietly capable. Underrated relative to its press footprint.