The Frontier: Where Agents Are Heading
Agents changed more between late 2025 and late 2026 than in any year before. This final module steps back to look at the trends that matter, the problems nobody has solved yet, and how to keep up once the course is over.
Long-horizon and always-on agents
The clearest trend is how long agents can work on their own. METR's time horizon, the length of task (in human-expert time) an agent finishes with 50% reliability, has been growing about 10x per year, doubling roughly every 3.5 to 4 months.
- Claude Opus 4.6 (February 2026) measured about 12 hours at 50% reliability.
- An early Claude Mythos Preview (April 2026) measured "likely at least 16 hours", which is beyond the range METR's current task suite can measure reliably.
- METR's time-horizons page is no longer actively updated, and there are no verified horizon figures for Claude 5.x, GPT-6 or Gemini 4. Measuring has become the bottleneck.
Longer horizons change the product shape. Instead of agents you chat with, vendors now ship agents that keep running, with their own computers, schedules and memory:
OpenAI "dots"
Announced at DevDay on 2026-09-29: always-on agents with their own cloud computer, running on GPT-6 Astra.
Claude Managed Agents
Launched April 2026 as a hosted harness with state, memory, permissions and scheduling, adding multi-agent orchestration and outcomes in beta in May 2026.
Bedrock AgentCore
AWS's managed runtime, where agents can run for up to 14 days.
Agents beyond coding
Claude Cowork (January 2026) brought the coding-agent harness to general knowledge work; browser agents such as Claude in Chrome moved to general availability.
Why it matters
An agent that runs for days needs everything from Building agents at full strength: durable state, sandboxes, budgets, approvals and monitoring. Capability has moved faster than the operational practices that make it safe.
Self-improving agents
Can an agent make itself better? In 2026 the answer is "partly, and mostly by improving the harness around a fixed model."
Darwin Gödel Machine
From Sakana AI and UBC, presented at ICLR 2026. A coding agent rewrites its own scaffold, tests each change empirically on coding benchmarks (SWE-bench and Polyglot), and keeps an open-ended archive of variants to build on.
AlphaEvolve
Google DeepMind, May 2025. An evolutionary coding agent that improves artifacts such as algorithms rather than itself. Models propose changes, automated evaluators score them, and the best survive.
Harness optimization
2026 papers on evolvable agent harnesses and memory evolution point the same way: keep the model frozen and evolve the prompts, tools and memory around it. That is the practical form of self-improvement today.
The catch: the evaluator is everything
Self-improvement only works when there is a trustworthy automatic score to climb. If the score can be gamed, the system will find the gaming, as the reward-hacking examples in Evaluating agents show. Good evaluators are the limiting resource.
The agent economy and protocol consolidation
In 2025, agent protocols looked like a format war. By late 2026 they had largely consolidated under one neutral home, the Linux Foundation's Agentic AI Foundation (AAIF).
- December 2025: AAIF formed, hosting Anthropic's MCP, Block's goose and OpenAI's AGENTS.md. Its platinum members include AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI.
- Late summer 2025: IBM's Agent Communication Protocol merged into A2A.
- 2026-07-28: MCP's largest revision so far, with a stateless core that lets servers run behind ordinary load balancers and on serverless platforms.
- 2026-08-27: A2A joined AAIF as a Growth Stage project, keeping its own steering committee.
The working model is MCP for vertical connections (agent to tools and data) and A2A for horizontal ones (agent to agent). On top of that, a commerce layer is forming: AP2, ACP, UCP, x402 and MPP all aim to let agents discover, pay and prove user consent. That layer has not consolidated yet. See Agent commerce for the details.
Open weights and multi-agent science
The open-weight frontier. Openly downloadable models are now serious agent engines, and Chinese labs lead this tier.
| Model | Released | Notes |
|---|---|---|
| GLM-5 (Z.ai) | February 2026 | 744B total parameters, 44B active |
| DeepSeek V4-Pro / V4-Flash | 2026-04-24 | MIT license |
| Kimi K3 (Moonshot) | 2026-07-16; weights 2026-07-27 | 2.8T parameters; modified-MIT license |
| Qwen 3.8-Max | 2026-08-03 | About 2.4T parameters; strong vendor-reported computer-use scores |
Open weights matter for agents because they let you run inside your own network, fine-tune on your own tasks and control costs on very long runs.
The science of multi-agent systems. Research has become more skeptical and more precise about when several agents beat one:
- The MAST study (UC Berkeley, NeurIPS 2025) annotated over 1,600 traces and found 14 failure modes. Most came from system design and coordination, not from weak base models.
- 2026 work such as "The Illusion of Multi-Agent Advantage" found that automatically designed multi-agent systems often fail to beat a single model using chain-of-thought and self-consistency at equal cost.
- Work on scaling through diversity suggests why: copies of the same agent make correlated mistakes, so adding more of them saturates quickly, while heterogeneous agents bring genuinely new information.
- Distillation work such as AgentArk explores folding multi-agent behavior back into a single agent.
Practical upshot
If you use multiple agents, make them different (different models, tools, context or roles) and give them independent subtasks. Ten clones of one agent arguing is mostly an expensive way to agree.
Open problems
Agents are impressive and still unreliable. These are the problems that will decide how far they go.
Reliability
The gap between 50% and 80% time horizons (about 12 hours versus about 1 hour for Opus 4.6) shows agents that can sometimes do long work but cannot yet do it consistently. pass^k, not pass@k, is the bar for production.
Security
Prompt injection is widely seen as an architectural property with no full fix. Containment patterns work but limit what agents can do. Making agents both capable and safe on untrusted input is unsolved.
Evaluation
Benchmarks saturate within months, get contaminated, and fork into versions. METR's measurements beyond about 16 hours are unreliable. We are building agents faster than we can measure them.
Cost
Agents use roughly 4x the tokens of chat and multi-agent systems about 15x. Prices are falling, but long-running, always-on agents multiply usage.
Memory
Agents still struggle to learn across sessions without accumulating junk or poisoned memories. Memory benchmarks exist, but vendor scores are self-reported and not comparable.
Identity
Who is this agent, who does it act for, and what is it allowed to do? NIST, IETF drafts and vendor products are working on answers, but none is a finished standard yet.
Timeline: key milestones, 2025 to 2026
| Date | Milestone |
|---|---|
| Mar 2025 | METR publishes its time-horizon method, finding a 7-month doubling time from 2019 to 2025 |
| Apr 2025 | Google announces A2A with 50+ partners; it moves to the Linux Foundation in June |
| May 2025 | Google DeepMind's AlphaEvolve; Claude Code reaches general availability |
| Jun 2025 | Simon Willison names the lethal trifecta |
| Sept 2025 | Claude Code SDK renamed the Claude Agent SDK; Google announces AP2 for agent payments |
| Oct 2025 | OpenAI AgentKit (DevDay); LangChain and LangGraph 1.0; Meta's Agents Rule of Two |
| Dec 2025 | Agentic AI Foundation formed; OWASP Top 10 for Agentic Applications; Agent Skills published as an open standard |
| Jan 2026 | MCP Apps launch; Google's Universal Commerce Protocol; Claude Cowork |
| Feb 2026 | Claude Opus 4.6 measured at about a 12-hour time horizon; NIST AI Agent Standards Initiative |
| Mar 2026 | Stripe and Tempo launch the Machine Payments Protocol |
| Apr 2026 | A2A passes 150 supporting organizations; Claude Managed Agents launches; DeepSeek V4 |
| Jun 2026 | Strands Agents 1.0 and AgentCore Harness GA; Pydantic AI 2.0 |
| Jul 2026 | Kimi K3; MCP 2026-07-28 stateless spec; EU Digital Omnibus defers high-risk AI Act deadlines |
| Aug 2026 | A2A joins the Agentic AI Foundation |
| Sept 2026 | GPT-6 Astra; Claude Opus 5.5 (66.4% on Terminal-Bench 4.0, vendor-reported); OpenAI DevDay introduces dots |
| Oct 2026 | Gemini 4 Argon broadly available; OpenAI Agents SDK 0.24.0 |
What to watch next, and how to keep learning
This field moves monthly. These are the threads most likely to matter over the next year:
- How we measure very long tasks, now that time-horizon suites top out around 16 hours.
- Adoption of MCP's stateless spec and A2A v1.0 signed Agent Cards across frameworks and platforms.
- Which commerce protocols win, and whether payments for agents consolidate the way tool and agent protocols did.
- Agent identity moving from NIST concept papers and IETF drafts toward real standards.
- The EU AI Act's high-risk obligations, now due from December 2027.
- Always-on agents from OpenAI, Anthropic, Google and AWS, and how users and businesses actually use them.
- 1
Follow primary sources
Read the MCP blog, the A2A blog, METR, the OWASP GenAI project and the engineering blogs of the major labs. Treat leaderboard aggregators and "top 10" posts with caution.
- 2
Read a few careful commentators
Simon Willison's blog is a reliable running record of agent security and tooling. Pick a small number of sources that show their evidence.
- 3
Build something small every month
Try a new SDK, protocol feature or model on a real task. Hands-on time teaches what release notes cannot.
- 4
Keep your own eval set
When a new model or framework arrives, run your evals rather than trusting headlines. Your tasks are the benchmark that matters to you.
Course wrap-up
You have covered the full arc of agentic AI in twelve modules. A single agent is a model in a loop with tools and a carefully managed context. MCP and A2A connect it to tools and to other agents. Multi-agent systems, commerce and computer use extend what it can do. Frameworks, evaluation and security turn it into something you can ship.
Start simple
Use the simplest design that works, and add agents, graphs or protocols only to fix a limit you can name.
Measure honestly
Build your own evals, run them on every change, and read transcripts, not just scores.
Contain by design
Assume agents will be tricked. Limit what a fooled agent can reach and require approval for what cannot be undone.
Stay curious and skeptical
Capabilities are growing fast. So is hype. Primary sources and your own experiments are the best guide.
Where to go from here
Revisit any module from the course overview, or put what you learned to work with ready-made agents on AIKind agents. Thanks for learning with us.
Key takeaways
- Agent time horizons have grown about 10x per year, and vendors now ship always-on agents with their own computers, schedules and memory.
- Self-improvement in practice means evolving the harness around a frozen model, and it is only as good as the evaluator it optimizes against.
- Agent protocols consolidated under the Agentic AI Foundation: MCP for tools, A2A for agents, with a commerce layer still forming.
- Open-weight models are now credible agent engines, and multi-agent research shows that diversity, not headcount, is what helps.
- Reliability, security, evaluation, cost, memory and identity remain open problems.
- Keep learning from primary sources, small experiments and your own evals.
Sources & further reading
- METR: Time Horizon 1.1
- LessWrong: METR time horizons now 10x/year
- Darwin Gödel Machine (arXiv 2505.22954)
- Self-Improvements in Modern Agentic Systems: A Survey (arXiv 2607.13104)
- TechCrunch: Linux Foundation launches the Agentic AI Foundation
- A2A joins the Agentic AI Foundation
- MCP 2026-07-28 specification release
- Decrypt: OpenAI DevDay 2026 announcements
- Why Do Multi-Agent LLM Systems Fail? (MAST)
- Understanding Agent Scaling via Diversity (arXiv 2602.03794)
- Simon Willison: Kimi K3