AI Agents in 2026: What Actually Shipped, What Flopped

Every product roadmap in late 2025 had the word "agent" on it, every keynote had an autonomous demo, and every deck promised the end of the SaaS form. Seven months in, the split is sharper than the slides suggested: some agents quietly took over real work, most of the loud ones did not. Here's the mid-year field report: what shipped, what flopped, what's still demo-ware, and what we expect to matter going into 2027.

What actually shipped and stuck

The winners share an unsexy profile: narrow scope, a clear definition of success, and a home inside an existing workflow rather than a replacement for it. Users don't notice the agent; they notice one annoying part of their job got faster.

• Coding agents and async PR bots. The clearest win of 2026. Claude Code, OpenAI's Codex, Cursor and GitHub Copilot's cloud agent run tasks in the background and open pull requests a human reviews. Cursor crossed roughly $2B in annual run-rate revenue by February.
• Customer support resolution. Intercom's Fin, Decagon, Sierra and Ada resolve or route a conversation end to end, increasingly billed per resolution rather than per seat. It works — read the fine print on the numbers.
• Meeting and email summarization. So commoditized nobody sells it as a product; it's a checkbox in the suite you already pay for.
• RAG-backed internal search. The least flashy shape here and the most useful. "What's our policy on X?" answered in seconds, with a citation.
• Calendar and scheduling. Not the "AI chief of staff" demos — the narrow version: find a time, draft the invite, send it.

None of these got a keynote slot. All of them shipped.

What flopped

The losers share the inverse profile: open-ended goals, fuzzy success criteria, no human in the path of anything important. Spectacular in a 90-second video, in pieces by week three.

• "Autonomous business agents." The AI-employee replacements for your inbox, your books, your hiring funnel. Gartner now expects more than 40% of agentic AI projects to be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. It also named the sales motion: agent washing, relabeling chatbots and RPA scripts as agents. Gartner reckons only about 130 of the thousands of self-described agentic vendors are the real thing.
• Browser and computer-use agents on the open web. OpenAI retired Operator in August 2025 after it struggled with checkout flows, CAPTCHAs and session state, folding the idea into Atlas. Google quietly folded Project Mariner into other products in May 2026. Computer-use agents still score in the high 30s on OSWorld, the standard benchmark for full desktop tasks — a fine research number, a terrible production number.
• Multi-agent showcase demos. Five agents debating each other to write a marketing plan. Great theater. Princeton research found a single agent matched or beat multi-agent setups on most benchmarked tasks given the same tools and context, with multi-agent buying roughly two points of accuracy for double the cost.
• Long-horizon planning agents. Anything that tracks a goal across days of state without a human nudge still drifts, no matter what the demo loop shows.

The tech isn't bad; the framing was wrong. "Be autonomous" is a far harder ask than "do this specific thing well, then stop" — we unpack that in how prompts, skills and workflows actually fit together.

The agents that stuck weren't the ones in the keynote. They were the ones nobody made a video about.

The plumbing that made the boring wins possible

2026 shipped real agents because the unglamorous infrastructure layer finally caught up. Three things did most of the work.

• MCP became a governed standard. Anthropic open-sourced the Model Context Protocol in late 2024 and donated it to the Agentic AI Foundation under the Linux Foundation in December 2025, turning a vendor project into something procurement will sign off on. The 2026-07-28 spec revision is the big one: a stateless core on ordinary HTTP infrastructure, a Tasks extension for long-running work, authorization aligned with OAuth and OpenID Connect, and a formal deprecation policy. We unpack the mechanics in MCP: the Model Context Protocol explained.
• The framework wars ended in consolidation. OpenAI's Agents SDK, Google's ADK, Anthropic's Claude Agent SDK and Microsoft's Agent Framework 1.0 (which merged AutoGen and Semantic Kernel into one supported SDK) all landed in the first half of 2026, and agent-to-agent messaging settled under the Linux Foundation as A2A. Fewer bets, less rework.
• Evals and observability grew up. Langfuse, LangSmith and Braintrust are normal line items, and OpenTelemetry's generative-AI conventions mean traces aren't locked to one vendor. That matters: most agent incidents are tool-call failures, truncated context and runaway loops, none of which a normal APM stack sees.

What hasn't changed

Three things look exactly like they did in 2024.

• Humans are still in the loop. Not as a transition phase, as the design. Every successful 2026 agent has a checkpoint where a human matters. The teams that removed the human wrote the postmortems.
• Evaluation is still hard. "Did the agent do a good job?" is mostly answered by a person reading outputs, backed by unit evals and sampled production traces. Tooling got better; it did not get easy.
• Fuzzy goals still fail. Models got smarter. They did not get more telepathic. An agent given "improve our onboarding" still doesn't know what you mean; "reduce day-7 drop-off on the email step" works. Gartner's read on the cancellation wave is basically this: models can't reliably chase business goals nobody bothered to define.

What to watch going into 2027

Everything below is a prediction, not a report. Take it as our read on where the leverage is, not as settled fact.

• Payments and identity get solved before autonomy does. Google's Agent Payments Protocol launched with 60-plus collaborators including Mastercard, PayPal and American Express, Coinbase's x402 moved under the Linux Foundation this July, and Stripe wired up stablecoin settlement earlier in the year. Our bet: 2027's agent fight is about authorization, liability and identity, not smarter planning.
• MCP's stateless revision quietly unlocks long-running agents. Once the Tasks extension and ordinary HTTP hosting are normal, "an agent that runs for two hours" stops being an architecture project. Expect a wave of scheduled and event-triggered agents.
• Agent washing gets priced out. After a cancellation wave, buyers start asking for eval numbers and trace access before signing. Good for anyone who built something real.
• Computer use goes API-first. We expect a continued retreat from "drive any website" toward integrations and narrow scripted paths, because the open web is adversarial and prompt injection through page content is unsolved.

The high-autonomy frontier hasn't gone away. It's just gotten more honest.

Where people go wrong (and when to call a pro)

If you're scoping an agent project, this is the short list to skip.

Without an in-house AI org, the starting point is almost never the framework. It's the workflow. Pick one narrow, painful step, define "success" in a sentence, build the smallest agent that moves that number. How to build your first AI agent covers it end to end, and our services page shows how we help teams ship one.

Frequently asked questions