AI Engineer Summit 2023
The Future of Work
Read the talk
From Chat to Completed Work: AutoGPT’s Agent Roadmap
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 Poots
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.
Connect agent development through a shared protocol
Work on the original open-source agent continues through capability improvements and the incorporation of research. Alongside that agent, the team introduces Forge, a standardized template intended to make it easier for other developers to build agents. The scope expands from developing one agent to supplying a starting point for other agent creators.
A development UI provides a way to interact with an agent and improve it iteratively. The displayed interface brings together connected task nodes, a SortCsv test card, and User and Agent panels displaying a JSON artifact. It makes the development workflow concrete: tasks, test controls, and an artifact can be inspected within the same interface.
These tools connect through Agent Protocol, which Silen attributes to the AI Engineer Foundation, alongside other industry standards. Its architectural role is a common interface between agents and the tools around them. An agent implementing that protocol can use the compatible benchmark, frontend development tool, and other offerings built on it. Compatibility depends on the shared contract rather than on every developer building the same agent.
The tools are still prerelease, but participants are already using them in a virtual hackathon with a reported $30,000 cash prize pool. That use supplies feedback, bug fixes, and insights for subsequent development. The hackathon therefore serves both as an invitation to build agents and as a source of practical experience with the tooling.
Suggest correction
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.
Read the complete timestamped transcript
- 0:00
[upbeat music] Thank you, San Francisco, for the warm welcome.
- 0:18
I'm Toran, the creator of AutoGPT, and I'm excited to show you all what the brilliant minds at AutoGPT have been working on over the past months. I'm going to hand off the stage now to Silen, one of our founding AI engineers.
- 0:35
Thank you, Toran. Thank you. There you go. Thank you, Toran. Um, I want to talk about something that I think not many of you realize. I didn't realize this for a long time.
- 0:48
We're not achieving the peak of our potentials.
- 0:51
We can all work faster, we can work better, and we can do more with less time and less t- and less stress.
- 0:59
Let's take this spreadsheet, for example. I don't know about you, but I've stared at this interface for hours on end, and, um, I'm sick of it. How would you go about filling out this spreadsheet, all right?
- 1:09
It's the lead generating. Name of the company, you got the links. Uh, what you'd probably do
- 1:16
is go on Google, you search, copy/paste, maybe go on LinkedIn, copy/paste, back to Google, over and over and over again for hours, going back to the same interface, going back to the same websites.
- 1:28
But what if instead of all that, you can just chat and you get the same end result, a filled out spreadsheet with all the leads? All right, let me give you another example.
- 1:39
We all have unread messages, right? Not because we're lazy, allegedly, but because we're overwhelmed.
- 1:47
Now, how would you go about cleaning out your inbox?
- 1:50
You'd sit there for hours and hours sending the same variation of the same email.
- 1:56
But what if you could just chat? Last example, I promise. Actually, uh, these emails will now be leads in your inbox, uh, instead of, uh, just unread emails. Last example.
- 2:12
Say you're a company or a developer. You spend millions of dollars developing apps that take weeks, months, sometimes even years,
- 2:22
sitting there copy/pasting anyways because you're probably using ChatGPT. I know I am copy/pasting. But what if instead of all that effort, you just chat?
- 2:33
I think you get the point. There's a reason you've heard of AutoGPT.
- 2:39
AutoGPT inspired the minds of millions. It gave hope to what a world could look like where we all reach our full potential. The light at the end of the tunnel.
- 2:49
You could see the sparks of digital artificial intelligence.
- 2:53
And in this world, everyone goes from using their minds mostly to execute menial tasks with only 10% of their brains being used for creative work
- 3:05
to becoming creative masterminds, orchestrating the peak potential of their lives. And in this world, we're all AI engineers, whether you know it or not.
- 3:16
And people have noticed AutoGPT was the fastest repository to 100,000 stars. Every major new- news network picked up on this.
- 3:26
Everyone understands what the potential of this is, and it kicked off a whole new field of development, a whole new paradigm of augmenting humans to give them time back and live a more stress-free life.
- 3:40
And even the major players in the space all realized how big of a deal it is and work-- and now work on these agents.
- 3:48
And so I want to hand off to the primary open source developer at AutoGPT to talk a, uh-- to talk a little bit more about the open source repo.
- 4:01
Thank you, Silen. Thank you, Silen, and hello, San Francisco. [audience applauding] [audience cheering]
- 4:13
I think all of us being here is a real testament to the power of open source. And on that note, we have some really exciting news to share.
- 4:24
Because just last week, our open source repo, AutoGPT, hit 150,000 stars on GitHub. [audience applauding] [audience cheering]
- 4:45
Of course, metrics are fun, but to me it is so much more than just a number.
- 4:52
It is the 150,000 people who took an interest in what we're doing and decided to click that button. So if you star our repo, then thank you.
- 5:04
It is also the 460 plus contributors who took the-- their time and effort submitting thousands of pull requests and issues
- 5:15
in the process, and to all of them as well, thank you so much.
- 5:21
It is also the 47,000 members of our online community
- 5:28
and all the interesting and insightful interactions that they've given us.
- 5:33
It's been a wild ride at times, but it, it has allowed us to do and learn so much in the past six months.
- 5:43
And I'm extremely excited for what is to come based on that.
- 5:52
Now, I've already said it, but we could not have done this without our community, and community matters.
- 5:59
So we are committed to fostering, to growing, and to empower this community
- 6:07
And to build the future together. And I'll hand it back to Silen to tell you what that means
- 6:19
Thank you, Poots. And we haven't stayed stagnant since the open source agent originally came out. We've continued to work on it, and we've continued to improve its capabilities and implement the latest cuttin- cutting-edge research.
- 6:30
But we've also been working on some other things. To show our commitment to the agent space and the open source ecosystem, we built a forge, which is a template for any agent, for any agent creator to have a better time to develop their agents with a standardized template.
- 6:46
We also built a dev tool UI to easily interact with and iterative- iteratively improve your agent using an intuitive interface.
- 6:54
All of these tools are built on top of the Agent Protocol from the AI Engineer Foundation and other industry standards to maximize compatibility and interoperability. Anyone who implements this protocol can use our benchmark, frontend dev tool, and other offerings built on top of this protocol.
- 7:11
And while this dev tool template is in beta or in alpha, uh, it has served, uh, our participants of the current hackathon we're running, where we have $30,000 in cash on the line.
- 7:21
And we've learned a lot from this. We've received a lot of great feedback. We've received a lot of bug fixes, uh, and insights that we're going to take into the future.
- 7:29
One of those insights is that code is king.
- 7:35
We've realized that coding agents are the fundamental agents, uh, of the world.
- 7:42
Let me tell you, the digital fabric, the fundamental digital fabric is code.
- 7:49
Our goal is to build a general a-- generalist agent, yes, but code is the s- is the stepping stone to AGI.
- 7:56
A motivated coder can get anything done, uh, except forget a bed frame. [laughs]
- 8:04
Another thing that we've learned over time is that without a compass, you don't know where you're going. You know, at the start of the repo where we're getting thousands of pull requests, and
- 8:14
that's a pull request every two hours. We had no way to know whether the pull requests were good and how do we even test these pull requests. We didn't have a real direction.
- 8:24
It took time to test these, and it was unnecessarily costly. And so we created a compass. We created a benchmark to direct the development of the open source repo and quantitatively know if we were improving.
- 8:36
It's an easy way to know if your agents are improving down different categories, and people are currently benefiting from this for the virtual hackathon.
- 8:44
And this is just cool. We've been running this in our CI pipeline for the past couple months on different open source, uh, agents within the ecosystem. And what the tests have shown is that we're on the brink of something special.
- 8:58
These agents have showed continual improvement, and don't worry, I wouldn't put this in a research paper. The-- It's very noisy. It's very messy. But there is a continuous trend from thirty-five to fifty-five percent.
- 9:08
Uh, this is just a graph of the success rate on the benchmark over time, over the month of August.
- 9:15
Another thing that we're committed to is safety.
- 9:20
As the ecosystem grows and as the capabilities of agents increase,
- 9:25
there's always questions of trust and reliability, and these are problems that AutoGPT is committed to. One of these problems is prompt injection, which will always be there. OWASP, uh, one of the big security organizations, ha- has talked about this and said this is one of the big problems that not just language models face, but also agents.
- 9:44
It's essentially when agents visit a website, which all agents need to do, and, uh, the website has something malicious. Um, and then the LLM is like, "All right, I need to be doing that now."
- 9:55
And you can see that there, uh, in this example.
- 9:59
Then there's this other, uh, category that I like to call, uh, innocently malicious, where, uh, agents are just bad sometimes. It's the truth. And in this example behind me, uh, this person asked, uh, an open source agent to delete all the JSON files within a directory, a specific one, and the agent ended up deleting all the JSON
- 10:21
files on a laptop. And this is gonna continue to be a problem. If we want agents to do the things that humans can do, they will need root access.
- 10:30
And so within AutoGPT, we're committed to and think about these problems extensively, and we've been working on an, a research paper to solve some of these issues. Um, and in order for agents to be commercially viable and trusted, these safety problems need to be solved.
- 10:45
You can't have a ninety-nine percent success rate. It has to be one hundred percent. That one email that's sent could be a lost contract or a lost lead. And so this is fundamental, not just to the development of the open source agent, but to all agents out there.
- 11:00
After all, our end goal is a digital AGI to augment all of humanity. And I'm gonna invite Craig to announce some exciting news regarding some developments with AutoGPT.
- 11:17
Hello. So it's been a wild journey from zero to here in six months,
- 11:27
and we keep stressing this because it's so important to us. We're only here because of our community, because of that shared passion in pushing the frontier of what AI agents can do.
- 11:40
So we're really excited to announce that Redpoint Ventures has invested twelve million dollars in turning this vision to a reality. [audience applauding]
- 11:56
Now, this isn't just funding. This is that, uh, them showing their deep belief in our mission and their dedication to open source. That's why we went with them, because they are so dedicated to staying open source, and that's really important to every single one of us working on this project.
- 12:17
Now, this is where we need you. With this funding, we wanna grow our team and add more passionate individuals.
- 12:26
So join us. Message us. Join our Discord community, and let's all help make the world's best open source generalist agent. Together, we can redefine the future of work. [audience applauding] [upbeat music]