Q2 2026 was the quarter the AI agent narrative caught up to reality. Less hype, more shipping. Here's what mattered.
The model layer — incremental gains, no breakthroughs
The big three model families shipped reliable improvements:
- Claude Sonnet 4.6 — better coding, fewer refusals, faster
- GPT-5 went through 3 stability updates — production-grade now
- Gemini 2.5 Pro — 1M context standard, Deep Research improvements
No GPT-5.5 / Claude 5 / Gemini 3 announcements. The market's reading of "the model layer is mostly flat for 6-12 months" looks correct.
What this means for agents: capability is no longer the bottleneck. Distribution and integration are.
The agent layer — three categories matured
Coding agents
Cursor Agent v2 shipped in March — multi-file edits got genuinely reliable. Devin reduced "spiral" failures (the agent burning hours on dead ends) by ~50%. Claude Code added MCP-server integration.
The takeaway: developers using AI agents in 2026 ship ~30-40% faster on well-scoped work. The reliability tier of these tools genuinely moved from "experimental" to "production daily driver" this quarter.
Sales agents
Artisan Ava, 11x, Clay all shipped meaningful upgrades. The big shift: research depth replaced volume as the differentiator. Generic AI cold outreach (templates with mail-merge) dies; specifically-researched AI cold outreach (5+ min per prospect, real signal-based hooks) gets 10-15% reply rates.
The takeaway: AI SDRs went from "marginal cost reduction" to "actually beating human SDRs at certain motions". $300-500/mo tool budgets are replacing $80-120k/year hires for outbound-only roles.
Voice agents in customer service
This is the breakthrough of Q2. Sierra, Decagon, Parloa, Vapi all crossed the "convincing enough that callers don't ask for a human" threshold. Hold music drops, deflection rates hit 50-60% in production, and the gap between AI and human CS reps closed visibly.
The takeaway: customer service is the highest-volume real-world AI deployment in 2026. Companies handling 10k+ tickets/month are saving $500k-2M/year. The economics finally work.
What died this quarter
- Generic "autonomous workflow" platforms without category focus. Most pivoted to vertical agents or shut down.
- "AI for X" wrappers with no model differentiation. Customers consolidated to direct ChatGPT/Claude subscriptions + tools.
- Multi-agent orchestration platforms without clear use case. The marketing was bigger than the demand.
The pricing story
Subscription fatigue is real. Customers paying $200+/mo across 5-7 AI tools started consolidating in Q2. Winners:
- Tools with clear, defendable value (Cursor, Claude, Midjourney)
- Tools that replace expensive humans (voice CS, AI SDRs)
- All-in-one platforms with real depth (Lindy, Notion AI)
Losers:
- Tools at $50-100/mo that overlap heavily with $20/mo ChatGPT/Claude
Expect continued price compression on agent middleware in H2 2026.
What we got wrong in Q1
A confessional section. Things we said in Q1 that didn't pan out:
- We bet Manus would dominate browser agents — it grew but other browser agents closed the gap faster than expected
- We were skeptical Voice CS would scale — wrong, it scaled fast
- We over-rated multi-agent platforms — most underdelivered
We'll keep tracking. Next quarterly post in Q3.
Where the puck is going
Watch for:
- Voice agents expanding beyond CS — sales (Vapi), interviewing, internal IT helpdesks
- Code agents on longer horizons — 4-8 hour unattended runs becoming reliable
- Multimodal default — text + voice + image in one workflow as baseline, not premium
- Open-source agent stacks competing seriously with commercial — pressure on $50-200/mo middleware
For the underlying catalog see our agents index (88 indexed) and AI tools index (165 indexed).