Skip to main content

Start building authenticated and predictable agents with Portia AI

· 4 min read
Emma Burrows
Co-founder and CTO
Mounir Mouawad
Co-founder and CEO

Tired of your AI agents going off the rails (↗)? Well look no further 😅! We are releasing an open source developer framework that allows you to build agents that pre-express their actions, share their progress and can be interrupted by a human.

alt text

Portia AI was born from our tinkering together with Fintech co-pilots (closer to our home turf). We both believed that AI agents represented a paradigm shift in how software interacts with users and their environment. Amongst other major changes in the past year, AI has become a primary user interaction layer as evidenced by the overwhelming focus of traditional players in the space (e.g. Salesforce’s Agentic 2.0 press release (↗)), the shift from “Buy” to “Build” in the SaaS space is accelerating as folks like Klarna leverage AI to automate across functions (↗), and the rise of web agents and multimodal models.

The problem space​

We were inspired by the explosion of AI-powered use cases, but the challenges we encountered as we tinkered were also sobering. Through this experience and conversations with other developers we honed in on the following pain points:

  • Planning: Many use cases require visibility into the LLM’s reasoning, particularly for complex tasks requiring multiple steps and tools. LLMs also struggle picking the right tools as their tool set grows: a recurring limitation for production deployments.
  • Execution: Tracking an LLM’s progress mid-task is difficult, making it harder to intervene when guidance is needed. This is especially critical for enforcing company policies or correcting hallucinations (hello, missing arguments in tool calls!).
  • Authentication: Existing solutions often disrupt the user experience with cumbersome authentication flows or require pre-emptive, full access to every tool—an approach that doesn’t scale for multi-agent assistants.

Our proposed solution​

While AI engineers with deep expertise have been hacking their way through these issues, we wanted to democratise the solutions for all developers with Portia AI. As a first step, we are offering an open source Github repo (↗), augmented with elective cloud-hosted features to help speed up deployments, and accessible from the Portia dashboard (↗).

  1. Pre-expressed plans: Our open source planning agent guides your LLM to produce an explicit Plan in response to a prompt, weaving the relevant tools, inputs, and outputs for every step.
  2. Stateful, controllable agents: Portia will spin up a PlanRun and a series of execution agents to implement the generated plans and track the run state throughout execution. Using our Clarification abstraction you can define points where you want to take control of plan runs e.g. to resolve missing information or multiple choice decisions. Portia serialises the PlanRun state, and you can manage its storage / retrieval yourself or use our cloud offering for simplicity.
  3. Extensible, authenticated tool calling: Bring your own tools on our extensible Tool abstraction, or use our growing plug and play authenticated tool library, which will include a number of popular SaaS providers over time (Google, Slack, Zendesk, Github etc.). All Portia tools feature just-in-time authentication with token refresh, offering security without compromising on user experience.
Intrigued?

Give us a try on our live playground on our website (↗).

It’s early days–for us and for the ecosystem at large. Everything from LLM reasoning, to authentication and APIs in the age of AI agents is evolving rapidly. With Portia AI, we want to help developers ride this wave of innovation by combining intelligence, autonomy, and security.
If this resonates with you, let’s connect on our Discord channel (↗). We’re building and iterating based on feedback from our community, and we’d love to hear your thoughts. Together, let’s tackle the gnarly challenges standing in the way of the agentic future.

Emma & Mounir

Join the conversation​