AI Engineer Summit 2023
The Future of Work
Loading slides…
Read the talk
The Future of Work
AutoGPT’s roadmap connects conversational task requests to shared development tools, measurable progress, and the safety required for agents that act on a user’s behalf.
From a talk by Toran Bruce Richards, Silen Naihin and Reinier van der Leer
Ask for the completed spreadsheet
How do you fill a lead-generation spreadsheet when you start with company names and links? Silen Naihin’s opening example follows a familiar loop: search Google, copy a result, paste it into the sheet, visit LinkedIn, copy again, and return to Google. The work repeatedly crosses the same interfaces and websites to assemble the information one row at a time.
The proposed conversational alternative keeps the output unchanged: a completed spreadsheet of leads. Instead of directing every search and paste, the user asks for the result through chat. The intended shift is from operating each interface to specifying the work that should be completed.
The same proposal extends to two other workflows, in sequence:
- Inbox work: An overloaded inbox leads to hours spent sending variations of the same email. A conversational agent would handle that repetition, with the proposed business outcome of turning unread messages into leads.
- Application development: Silen invokes projects costing millions of dollars and taking weeks, months, or years, while developers still copy and paste from ChatGPT. Here, too, the proposed alternative is to request the work through chat.
These are motivating scenarios for conversational automation, not demonstrated implementations.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From creative orchestration to a development community
AutoGPT gives these scenarios a broader ambition: agents execute routine digital work so that people can spend more of their effort directing creative work. In this vision, being an AI engineer becomes less about personally implementing every step and more about orchestrating what an agent should accomplish. It is a proposal for human augmentation, with more room for creative decisions and less repetitive execution.
Silen describes AutoGPT as the fastest repository to reach 100,000 GitHub stars. He connects that attention to widespread news coverage and growing interest in agents from major technology companies. Those observations concern enthusiasm for the approach; they do not establish how reliably an agent can complete the opening tasks.
The team then describes the community behind the repository. Its reported figures distinguish people registering interest from people contributing work and feedback:
| Participation | Reported scale at the talk | Contribution to the project |
|---|---|---|
| GitHub stars | 150,000, reached the week before the talk | An expression of interest |
| Contributors | More than 460 | Thousands of pull requests and issues |
| Online community | 47,000 members | Conversations and insights |
The distinction matters: stars record attention, while submitted changes, issue reports, and discussions give the project material to build with and learn from.
The team credits those contributions and conversations with what it had built and learned over its first six months. Its commitment is to keep growing and empowering that community, making participation part of how the future agent will be developed.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Why prioritize coding agents?
One lesson from this work is, in Silen’s words, “code is king.” The reasoning starts with the environment an agent must act within: code is the underlying fabric of digital systems. Coding ability is therefore positioned as a foundation for accomplishing many different digital tasks. The destination remains a generalist agent; the team’s strategic thesis is that coding agents provide a stepping stone toward artificial general intelligence, or AGI. This is the rationale for prioritizing coding, not a demonstrated result about reaching AGI.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Use a benchmark as a development compass
A busy repository creates an evaluation problem before it guarantees progress. Silen recalls thousands of early pull requests and a cadence of one pull request every two hours. The team lacked a clear way to tell whether a proposed change was good or how to test it. Evaluating changes consumed time and money without providing a reliable direction for development.
The response was to build a benchmark that could indicate whether the agent was improving. Category-level results make it possible to examine different kinds of capability rather than relying only on a general impression. Hackathon participants use the same evaluation tool. The team also reports running benchmark tests in its continuous integration pipeline across different open-source agents for the preceding couple of months.
The plotted benchmark success rate trends from about 35% to 55% over August, according to Silen, who calls the results noisy, messy, and unsuitable for a research paper. The chart preserves both observations: jagged individual series and a rising dashed trend. It is presented as a signal for directing development, not a general success rate that can be assumed for agents on other tasks.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Two ways an agent can cause harm
As agents gain capabilities, trust and reliability become more consequential. The first failure mode is prompt injection, which Silen identifies through OWASP’s attention to the problem. An agent visits a website and encounters malicious content. The language model treats that content as something it should obey, redirecting its behavior. The dangerous transition is from reading external material as data to accepting it as instructions.
The next failure does not require malicious instructions. Silen calls this category “innocently malicious”: an agent causes harm while attempting a legitimate task. In the displayed incident report, a user asked an open-source agent to delete JSON files within one specific directory. The reported result was deletion of JSON files across the laptop. The requested operation was deletion in both cases; the consequential error was expanding its scope beyond the requested folder.
That incident leads into the question of permissions. Silen argues that agents able to do what humans can do will need root access. This is his position on the access required for broad autonomy. In that vision, increasingly powerful access makes mistakes more consequential, so capability development cannot be separated from safety.
The team reports working on safety research and a paper addressing some of these problems. No concrete solution is presented here. The commercial concern is whether users can trust an agent with actions whose consequences extend beyond the conversation.
Silen sets 100% rather than 99% reliability as the requirement for commercial use; he is stating a target, not reporting an achieved result. One wrongly sent email could cost a contract or a lead. That example makes the stakes specific: even a seemingly small action can have a business consequence that an overall success rate obscures. Safety is therefore part of the team’s stated goal of building digital AGI that augments humanity.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Fund the team behind the generalist agent
The closing announcement returns to the community and its shared interest in advancing agent capabilities. The team announces a $12 million investment from Redpoint Ventures. It presents the investment as support for turning the agent vision into a working reality.
The team says Redpoint’s commitment to remaining open source was a reason for choosing the investor. The intended next step is to grow the team, with an invitation to contact its members and join the Discord community. The closing objective remains ambitious and explicit: bring more people into building what the team wants to become the world’s best open-source generalist agent.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Source code and setup guidance for AutoGPT, with the original Classic agent separated from the later platform.
A common agent interface for creating tasks, executing steps, exchanging artifacts, and connecting compatible evaluation and development tools.
Further reading
Silen Naihin and colleagues propose monitoring agent actions to constrain unsafe behavior during tests on the open internet.
- The State of AI EngineeringArticle
An account of the inaugural AI Engineer Summit, including AutoGPT’s announced $12 million investment from Redpoint Ventures.
Full transcript
Loading transcript…