What an agentic development platform is and what it is not
Published · updated
An agentic development platform is the system in which coding agents write most of the code while the team writes the specification, sets the limits of the agents' autonomy and decides what reaches production. It covers the whole life of the software, from the context the agents read to the maintenance of what they built, and every session is logged, with its authorship.
On February 2, 2025, Andrej Karpathy described a way of programming he called "vibe coding," where you "forget that the code even exists." He accepted every change and no longer read the diffs, and he was clear about where it belonged: it was "not too bad for throwaway weekend projects." On February 4, 2026, he wrote that programming through LLM agents "is increasingly becoming a default workflow for professionals, except with more oversight and scrutiny," and named his current favorite term for that version: agentic engineering.
An agentic development platform is what that second way of working needs at company scale: many people, systems that cannot stop, auditors who ask who approved a change. The phrase is also used for developer tools. Google launched Antigravity in November 2025 as "our new agentic development platform." This page uses the term for the system around the whole software cycle and looks at where people sit in it.
What is an agentic development platform
Gartner calls the market around it Enterprise AI Coding Agents. In a May 2026 press release on that market, the firm described "a transformative shift from AI-assisted development to agentic software development, spanning the SDLC from planning to creating to reviewing code." Ithaca uses the label as vocabulary only. It is not an endorsement, and Ithaca makes no claim to appear in any Gartner report. For the lifecycle labels people use for the same shift, from AI-native SDLC to the AWS AI-DLC methodology, see AI-native SDLC, agentic SDLC and AWS AI-DLC: what each term means.
A coding agent is the part that writes. The platform is what makes that writing usable inside a company, and it holds four parts together:
- Approved context. A Knowledge base with the codebase, documentation and anonymized data, validated by the team, feeds each new session.
- Execution by coding agents. The agents write most of the code, under rules and limits the team sets.
- A session record. The request, the steps, the model, the cost, the result and the authorship of each session stay available for audit and improvement.
- Human supervision before production. People on the team read what goes to production. Level 5, the dark factory, is out of scope.
Take any one of them away and the result is something else: an assistant in the editor, a vendor handoff or a black box. With all four in place, a company can say who approved what reached production and reconstruct the session that produced it. Short definitions of coding agent, session record and the other terms on this page are in the glossary of software development with coding agents.
Agentic engineering vs. vibe coding
About six weeks after Karpathy's first post, Simon Willison drew the line in plain terms. Vibe coding, in his definition, is "building software with an LLM without reviewing the code it writes." His golden rule for production work is that he won't commit code to his repository "if I couldn't explain exactly what it does to somebody else." When an LLM wrote the code and a person then reviewed it, tested it thoroughly and can explain how it works, he writes, "that's not vibe coding, it's software development."
What separates the two practices is what happens after the agent writes:
| Vibe coding | Agentic engineering | |
|---|---|---|
| Who writes the code | The model | Coding agents |
| Who reads it | Nobody ("I don't read the diffs anymore," in Karpathy's words) | A person accountable for the result |
| Where it fits | Prototypes, demos, throwaway projects | Software that has to keep running |
| What is left afterward | The code | The code, the tests, the rationale and the session record |
| What the human does | Asks and accepts | Specifies, sets limits, checks and approves |
Willison also defends vibe coding for what it does well, such as low-stakes prototypes and personal tools, so the table is about scope rather than a verdict. An enterprise system with years of history, regulated data and customers on the other end falls outside that scope. The insight on agentic engineering vs. vibe coding covers when each one fits and how to take a prototype to production.
In Karpathy's description, "agentic" means you are not writing the code directly most of the time: "you are orchestrating agents who do and acting as oversight." A platform turns that oversight into something a company can repeat, audit and hand over, at defined points and with a record that outlives the session.
What an agentic development platform is not
| AI assistant in the IDE | Development agency | Dark factory (level 5) | Agentic development platform (level 4) | |
|---|---|---|---|---|
| Typical role | Speeds up the person typing | Delivers a project and hands it over | Specs in, software out | Coding agents execute; the team stays in control |
| Who writes most of the code | A person, with help | The allocated team | Agents, with no human reader | Coding agents |
| Who reads before production | Usually the author | Depends on the contract | Nobody, by design | People on the client's team, every time |
| What is recorded | The diff or pull request, at best | Tickets and pull requests | Little that can be audited | Request, steps, model, cost, result, authorship |
| After go-live | The tool stays in the editor | The engagement ends at handover | Operation without a clear trail | Sustainment, with the team supervising |
All four columns can involve AI. They differ in where autonomy stops, who reads, and what is left on record.
Agentic IDEs take the first column much further. Google describes Antigravity's Manager surface as "a mission control for spawning, orchestrating, and observing multiple agents across multiple workspaces in parallel," and frames the product as an evolution of the IDE. The category on this page answers a different question: how a company runs coding agents across the life of a system, with approved context, limits per stage, a record per session and maintenance after go-live.
Human in the loop: where the people sit
On Dan Shapiro's scale of coding autonomy, level 3 is where "you are the human in the loop": "You spend your days reviewing code. So much code. Your life is diffs." Almost everyone, he writes, tops out there.
Asking a person to watch a steady stream of output for the rare problem is an old difficulty. In "Ironies of Automation" (1983), Lisanne Bainbridge of University College London argued that "the more advanced a control system is, so the more crucial may be the contribution of the human operator." Citing vigilance studies, she noted that even a highly motivated person cannot keep effective visual attention on a source where very little happens for more than about half an hour. Her paper was about process plants and flight decks; the parallel with watching agent output is our reading.
The design question, then, is where a person decides rather than monitors. Work on coding agents points to three places.
Before the code, at the plan. Atlassian built HULA, a human-in-the-loop agent framework, into Jira for internal use and published the results with researchers from Monash University and the University of Melbourne (ICSE SEIP 2025). An engineer sets up the task, an agent proposes which files to change and how, and the engineer reviews, edits or regenerates that plan before any code exists. Across 663 real Jira issues, the agent produced plans for 527 and engineers approved 433. Code was generated for 376 issues, pull requests were raised for 95, and 56 were merged. Engineers reviewed the work at the plan and again at the code, and most issues stopped before a merge.
At the merge, with evidence. A research roadmap on agentic software engineering by Ahmed E. Hassan and colleagues at Queen's University and other institutions (a preprint first posted in 2025 and revised in 2026) argues that "rather than reviewing dozens of raw pull requests, a human reviewer should focus on auditing a structured Merge-Readiness Pack." The pack carries evidence of functional completeness, sound verification, engineering hygiene, a clear rationale and a full audit trail. Citing earlier research on automation, the authors note that higher levels of automation often shift risk from execution failure to supervision failure.
Around the delivery system, with controls. Google Cloud's 2025 DORA report, based on survey responses from nearly 5,000 technology professionals, concludes that AI "amplifies what's already there." It found AI adoption associated with higher software delivery throughput and, as in the previous year, with higher delivery instability. The report's reading is that without strong automated testing, mature version control and fast feedback loops, more change leads to instability.
On an agentic development platform, the team writes the specification and decides what the agents may do without prior approval and what needs a person's sign-off first. People read what goes to production at the merge, rather than watching every step as it happens. Every session is logged, with its authorship, so a review can start from the record of what the agent did. None of this replaces testing and fast feedback in the delivery system itself, which is where DORA locates the problem.
How Ithaca runs the platform at level 4
Ithaca operates an agentic development platform at level 4 of the autonomy scale Dan Shapiro proposed. In Shapiro's description, level 4 is the spec rung: "You write a spec. You argue with it about the spec." You review plans, then "leave for 12 hours, and check to see if the tests pass." His post does not say who reads the code at that level, and its illustration for level 4 has the human trusting the system's self-checks and checking the final results much later. Ithaca's practice adds a rule: coding agents execute, people on the team read what goes to production, and every session is logged, with its authorship. Level 5, which Shapiro describes as "a black box that turns specs into software," is the dark factory, and the platform stops before it.
The same four stages apply to a new product and to a system already in production:
- Knowledge base. Codebase, documentation and anonymized data go in before any new line is written, and the context the team approves feeds the next session.
- Build and refactor. Coding agents write most of the code. The team defines the specification, the rules and what needs human approval before it runs.
- Observability. Each session records the request, the steps, the model, the cost, the result and the authorship. Without that record, the gain cannot be audited or improved.
- Sustainment. Tickets come in and fixes go out, with the team supervising. Sustainment is a stage of the platform from the start.
The rules and limits the agents work under are a discipline of their own, described on the page about harness engineering for coding agents. The same structure applies to legacy modernization without stopping the business.
The platform does not remove risk. A failure is still possible; what changes is how much of the system it can reach and how quickly its cause turns up in the session record.
Dr. Consulta on the platform
Dr. Consulta modernized its subscription platform, a system with years in production, using the agentic development platform, with its own team supervising. Compared with the period before the modernization, infrastructure cost is 75% lower, and one feature that used to take about two weeks took up to three days. The feature is a representative example, not a measurement of the whole backlog.
Frequently asked questions
What is an agentic development platform?
An agentic development platform is the system that organizes a software cycle in which coding agents write most of the code, while the team sets the specification and the limits, reads what goes to production and keeps a record of every session. Gartner calls the surrounding market Enterprise AI Coding Agents; the label is vocabulary, not an endorsement of Ithaca.
What is the difference between agentic engineering and vibe coding?
Vibe coding, as Andrej Karpathy described it in February 2025, means accepting what the model writes without reading the diffs, which suits throwaway projects. Agentic engineering, his preferred term a year later, keeps agents writing the code and adds "more oversight and scrutiny." Simon Willison draws the line at review: code a person has reviewed, tested and can explain is software development.
What does "human in the loop" mean on an agentic development platform?
On an agentic development platform, human in the loop means people decide at the points that carry weight: the specification, what the agents may do without prior approval, and what goes to production. Every session is logged, with its authorship, so a reviewer works from the record of what the agent did and not only from the final diff.
Is Google Antigravity an agentic development platform?
Google uses the phrase for Antigravity, which it launched in November 2025 as an agent-first evolution of the IDE, where a developer can spawn, orchestrate and observe several agents. This page uses the term for the system a company runs around coding agents: approved context, limits per stage, a record per session, human review before production and sustainment after go-live.
Is this a dark factory?
No. A dark factory is level 5 on Dan Shapiro's scale: code nobody reads before it goes into production. The platform runs at level 4, where coding agents execute, humans supervise and the team sets the limit.
Where does the platform run, and does our code leave the network?
The platform is deployed in the client's environment, on a cloud account the client controls or in its own data center, and the repository is not sent to an Ithaca SaaS. Model calls follow the client's policy for provider, region and contract; during inference, code excerpts may go to the chosen model provider under that policy.
Book a demo
The next step is a demo of the agentic development platform on one part of the system that can be checked end to end. The deployment starts there, alongside your team, at level 4. Your team supervises, and every session is logged, with its authorship.
