AgentGPT

AgentGPT is Reworkd’s browser-based autonomous-agent project from the early AutoGPT wave. A user names an agent, gives it a goal and watches it propose tasks, execute iterations and feed results back into later steps. It made autonomous planning easy to understand without a terminal and remains historically useful for demonstrations.

The critical current fact is maintenance status: Reworkd archived the public repository on January 28, 2026, making it read-only. The hosted site and plan page were still reachable when reviewed, but reachability is not evidence of active development, current models, security patches or reliable support. Treat AgentGPT as a legacy educational system, not a default production platform.

What the loop actually does

 user goal
    |
    v
 generate task list
    |
    v
 choose next task ---> call model / search / plugin
    |                         |
    |                         v
    '------ add result to memory/context
                              |
                    done? ----+---- no --> another loop
                      |
                     yes
                      v
                final narrative

 More loops can amplify progress, error, cost—or all three.

The model interprets the goal, invents intermediate tasks and evaluates its own progress. Without an external verifier, it can convert a wrong assumption into several confident steps. Search results can be stale or hostile; task lists can drift; and a final summary can hide unfinished work. Autonomous iteration is useful only when each step has bounded authority and observable success.

Current status at a glance

Current displayed plans—with a legacy warning

The live plan page reviewed July 26, 2026 showed a Free Trial with five demo agents per day using GPT-3.5-Turbo, limited plugins and limited web search. Pro was displayed at $40/month with 30 agents/day, GPT-3.5-Turbo 16K, GPT-4 access, 25 loops per agent, unlimited web search, plugins and priority support. Enterprise was custom.

Those model labels and limits reflect an older era. GPT-3.5 Turbo has been retired from many modern workflows, and “GPT-4” is not a precise current model ID. Before subscribing, confirm checkout, renewal, cancellation, refund, supported models, last service update and an actual support response. Do not infer ongoing product investment from an active payment page.

Good educational tasks versus bad operational tasks

Write goals with termination conditions

“Research the market and make us successful” has no observable completion state. A better goal is: “Find five publicly documented competitors in category X, capture official URL, current plan price and quoted feature evidence as of today, and return a table with unknown fields marked—not inferred. Stop after five or after eight search iterations.”

• Define output schema, source quality and freshness date.
• Set maximum loops, time and spending.
• State forbidden actions and data.
• Require uncertainty and missing fields.
• Specify who validates the result and what authoritative source wins.

Evidence-first review

Do not review only the final narrative. Inspect the task trace, URLs, dates and exact source claims. Open each cited page and distinguish source text from the agent’s inference. A search snippet is not enough; a third-party article cannot establish official pricing when a vendor page exists.

Why more loops do not guarantee a better answer

Each iteration consumes context and may summarize earlier details. Errors can compound, duplicated tasks can consume the loop cap, and self-evaluation tends to favor plausible completion. Track unique verified facts gained per loop. Stop when two iterations add no evidence, repeat an action or contradict a confirmed constraint.

If a task needs durable waiting, retries, approvals and external writes, use a maintained workflow runtime with idempotency and auditability. A conversational loop should not be mistaken for distributed transaction management.

Self-hosting architecture and maintenance burden

The archived repository’s setup includes a Next.js/TypeScript frontend, FastAPI backend, MySQL database, NextAuth, Prisma/SQLModel, LangChain and provider keys. Optional Serper and Replicate keys extend search/media behavior. The automatic setup scripts were built around dependencies current at the time; modern Node, Python, Docker, database and OpenAI APIs may have breaking changes.

Forking creates a new product ownership obligation. Assign maintainers, establish a vulnerability process, update dependencies continuously and document data flows. If the purpose is simply to run modern agents, starting from a maintained framework is usually cheaper.

Secrets and access boundaries

Store provider keys outside source control, use low spending limits and rotate them. Do not expose the local app or API publicly without authentication, TLS and rate limits. Agent output and plugin content are untrusted. Never give an archived general agent broad filesystem, email, payment, cloud or social-media authority.

For a workshop, use a separate project/account, synthetic data and disposable keys. Monitor token/search consumption. Destroy the environment afterward and revoke credentials.

Historical value and what modern systems changed

AgentGPT demonstrated that a goal–task–execute loop could be made legible to non-developers. Modern production agent systems add explicit graphs, durable checkpoints, tool schemas, permissions, tracing, evaluations, human approvals and failure recovery. The design lesson is still valuable: users need to see what the agent plans. The implementation is no longer a safe default.

A one-hour classroom evaluation

1. Give AgentGPT and a supervised chatbot the same bounded public-research goal.
2. Limit the autonomous run to five iterations.
3. Mark each task as useful, duplicate, unsupported or off-scope.
4. Verify every source and count correct unique facts.
5. Compare time, tokens, correction effort and final completeness.

The likely learning is not “autonomy always wins” but when decomposition helps and when supervision prevents drift. Preserve the run as educational evidence; do not publish unverified output.

Alternatives in 2026

Frequently asked questions

Is AgentGPT still maintained?

No. The official GitHub repository was archived January 28, 2026 and is read-only.

Does the hosted site still work?

It was reachable when reviewed, and the plan page displayed offers. Availability does not imply active development or long-term support.

What does the current plan page show?

A free trial and a $40/month Pro plan with legacy GPT-3.5/GPT-4 wording, plus custom Enterprise. Verify checkout before paying.

Can I self-host it?

The source and setup remain available, but old dependencies and model integrations require security and compatibility work.

Is it suitable for production?

Not as an unmodified default. Choose a maintained runtime with permissions, evaluations, observability and durable failure handling.

Does it have live web access?

The hosted plan advertises web-search capabilities; self-hosting historically used an optional Serper key. Verify sources and protect against prompt injection.

Why keep the directory page?

It documents an influential agent interface and helps users avoid mistaking a still-live URL for a maintained modern platform.

Primary sources

• Official archived AgentGPT repository and status
• Official historical releases
• Archived security policy
• Hosted AgentGPT application
• Current displayed hosted plans
• Official self-host Compose definition
• Official historical environment configuration
• OWASP prompt-injection guidance
• MIT AI Agent Index for modern agent-transparency context

Last reviewed July 26, 2026. The repository is archived. Re-check hosted availability, model IDs, pricing, billing and support before any use or payment.