Frontier agents are here. They just don't run where your code lives.
Frontier agents are here. They just don't run where your code lives.
On 2 December 2025, AWS put a name and a definition on something the industry had been circling for a year: the frontier agent. Autonomous. Massively scalable. Able to run for hours or days without anyone checking in. Ten months later, that definition describes the shape of the entire market.
The engineering is real and the enterprise demand behind it is real. But every frontier agent shipped so far shares two properties, and neither one leads the launch post. The first is where it runs. The second is what it runs against.
This piece is about both.
What is a frontier agent?
A frontier agent is an AI system with three properties: it works autonomously toward an outcome rather than a reply, it scales across many parallel workstreams, and it keeps running for hours or days without a human turn in the loop. That is AWS's framing, and it's a clean one.
The contrast with the assistant era is the whole point. An assistant waits for your next message. A frontier agent is handed an outcome and comes back when it's done — or when it needs a decision only a human can make.
AWS shipped four to prove it. Kiro, the spec-driven development agent. An AWS Security Agent. An AWS DevOps Agent, which AWS puts at 86% accuracy on root-cause identification. A FinOps agent in preview. Four different jobs, one architectural bet: give the agent an objective and a long leash.
That bet mostly paid off. Code generation stopped being the bottleneck somewhere in 2025, and the frontier-agent wave is the industry admitting it out loud.
Who has shipped frontier agents?
Three vendors, in order, over four months.
AWS went first in December 2025 — named the category, defined it, and shipped agents against it.
OpenAI Frontier followed in February 2026 with the control plane rather than the worker: agent teams that split work between them, connectors that carry business context into the agent, permissions and auditing for the people who have to sign off, and forward-deployed engineers to get it installed. That last detail tells you who the buyer is.
Perplexity took the pattern into the enterprise in March 2026 with Computer, pushing agent execution onto machines inside the org's own workflow.
Credit where it's owed: this is good engineering, shipped fast, against demand that is not manufactured. Enterprises genuinely want long-running autonomous work. Nobody in this list is selling vapour.
Can frontier agents run on-premise?
Not today. All three run in the vendor's cloud, on the vendor's model, and your source code goes along for the ride.
For a large share of teams that is not a preference problem, it's a procurement problem. Regulated industries where data residency is written into the contract. Government and defence contractors. Enterprises whose codebase is the product, not a means to it. Consultancies working under client NDAs that never contemplated a third-party inference provider. Anyone whose security review contains the phrase "no third-party code egress."
None of those teams will win that argument internally, and most won't try. They'll watch the frontier-agent wave from the other side of a firewall while their competitors compound.
This isn't a privacy lecture. It's an availability gap. The most capable autonomous development tooling ever built is currently unreachable for the teams with the most to protect.
What is spec-driven development, and why does autonomy need it?
Spec-driven development means the requirements — PRD, personas, user stories, acceptance criteria, architecture — exist as a versioned artifact that the build is generated from and continuously reconciled against. Not a prompt typed once and forgotten. A document with a history.
This matters more with autonomy, not less. An agent working unattended for two days multiplies whatever you pointed it at. Point it at something vague and you get two days of confident, well-formatted drift.
AWS clearly saw this coming — Kiro is spec-driven by design. But a spec that lives only inside the tool that generated it goes stale the first time a developer fixes something by hand. One-way pipelines always end the same way: the PRD says one thing, the UI does another, the code does a third, and nobody can tell you which one is true.
That drift is the real reason AI-generated applications stall at "impressive demo." Vibe coding ships prototypes. Spec-driven ships products.
What we built instead: Codalio Studio
Codalio Studio is a desktop software factory. It starts where the other tools stop — not at the prompt, at the scope.
Studio generates the full specification before anything gets built:
-
PRD, personas, and user stories
-
Acceptance criteria per story
-
Architecture, ERD, and UML
-
An approved task list — nothing executes until you sign off on it
-
All of it versioned and shareable, outside any single agent
Then it conducts the agents your team already picked: Claude Code, Codex, Cursor CLI, OpenCode, OpenHands. They execute in your terminal. Your code stays in your git, on your disk. Bring your own key — Claude, OpenAI, Gemini, Grok, DeepSeek, Mistral, Ollama — with local models on the roadmap. One honest caveat, stated plainly: only spec generation runs in our cloud.
The loop closes in both directions. Change the PRD and the code follows. Change the code and the blueprint updates. Nothing goes stale, which is the part that makes long-running autonomy survivable.
Humans still sign off. Plan mode gates the start. A QA reviewer agent runs per task, and a person approves last — the same shape as a pull request, because that shape works.
The open-source half of the argument is Codalio Blueprint: MIT-licensed, ten skills, free, and it runs inside Claude Code, Cursor, Codex, Antigravity, and Gemini today. We're also working with the OpenHands team, because orchestration only earns its name if it's honest about not owning the agents.
What this means for your team this quarter
Three moves, none of which require buying anything.
First, ask every agent vendor two questions in writing: where does the code execute, and where is it retained? Vague answers are answers.
Second, make the spec a versioned artifact that lives outside the tool that wrote it. If your requirements can't survive switching agents, they were never requirements.
Third, put the human approval gate between plan and build — not after the PR. Reviewing two days of autonomous output is not review, it's archaeology.
Writing code got cheap this year. Knowing what to build, in what order, under which constraints, did not. Codalio is the only AI platform that takes non-technical founders from idea to PRD to working prototype — without the rebuild cycles that come from skipping the spec.
Get the spec half now
If the gap this essay named is your gap — agents that can build, and nothing solid to point them at — the fix starts before any tooling decision, with the questions you ask about your own product.
Codalio Studio is launching soon — no date until it's one we'd stand behind. The waitlistgets early access first, and our Discord hears the date first.
References
-
AWS frontier agents announcement — aboutamazon.com
-
AWS frontier agents product page — aws.amazon.com/ai/frontier-agents
-
OpenAI Frontier — openai.com/business/frontier
-
Perplexity enterprise agent coverage — Axios, March 2026