Agentic Company OS update: new industry teams, improved onboarding, evidence uploads, and customer-ready deliverables
I shared this project here previously when it was mainly a governed multi-agent execution prototype.
Since then, I have continued developing Agentic Company OS into something closer to a platform where users can create and operate AI teams for different types of work.
The main workflow is:
- Connect an Anthropic or OpenAI account
- Create a project
- Select an industry-specific team
- Choose how independently the team should operate
- Give the team a directive
- Watch the agents create tasks, collaborate, review work, raise questions, and produce deliverables
One of the biggest changes is the introduction of different verticals and team presets.
The available teams now include:
- Software Product Team for planning and building software products, with roles such as Product Manager, Requirements Engineer, Software Architect, Developer, Designer, Security Engineer, and QA.
- Cybersecurity Team for security assessments, evidence analysis, code and dependency review, infrastructure review, vulnerability confirmation, risk triage, and report generation.
- Legal and Contract Review Team for reviewing contracts, identifying risky clauses, preparing redlines, writing negotiation points, and producing a final risk memo for human approval.
- Marketing Team for market research, positioning, campaign planning, messaging, content creation, and review.
Each preset has its own:
- agent roles and responsibilities
- skills and permitted tools
- coordinator role
- task workflow
- review and approval structure
- expected deliverables
- model configuration
The goal is that selecting a different team should change more than the agents’ names. It should change how the project is decomposed, which tools can be used, who reviews the work, and what type of result is produced.
The platform now also supports custom LLM backends. In addition to Anthropic and OpenAI, users can connect Hugging Face Inference Endpoints, the Hugging Face serverless router, or another OpenAI-compatible endpoint. Different models can be assigned according to an agent’s role, allowing more capable reasoning models for coordinators and specialists while using faster or cheaper models for routine tasks. This makes it possible to combine commercial and open-weight models within the same agent team.
The cybersecurity workflow has received the most recent attention.
Users can upload evidence such as:
- logs
- configuration files
- scan reports
- dependency files
- source-code snippets
- existing security documentation
The cybersecurity agents can search and analyze this evidence while performing the assessment. The team can conduct static code analysis, secret detection, dependency and vulnerability checks, CVE/CWE research, infrastructure review, risk triage, and report preparation.
I have also worked on making the outputs more useful outside the application.
Projects can now produce structured deliverables that move through draft, review, approval, delivery, and customer acceptance. Reports can be exported as real PDFs, shared through a customer-facing portal, and accepted or rejected by the recipient.
Another major change is the onboarding experience. The application now guides a new user through five steps:
- Set up the LLM
- Pick a team
- Start the project
- Give the team a directive
- Watch the tasks progress
The dashboard adapts to the current stage instead of showing the entire operations interface immediately.
I have also been removing simulated tool results. A tool should now either perform real work or clearly report that the required integration is unavailable. The agents should not claim that they scanned a dependency, created a document, or inspected a file when that action did not really happen.
The larger idea behind the project is still the same: I am not trying to build another single-agent chat interface.
I want to explore what happens when AI work is organized more like a company:
- specialized teams for different industries
- explicit roles and responsibilities
- task delegation and collaboration
- review and quality-control stages
- human approvals and escalation
- controlled access to tools
- project memory and evidence
- customer-facing deliverables
I would especially appreciate feedback on these questions:
- Which vertical would you actually use?
- Are the current team presets specific enough?
- Which team should I build next?
- Would you use this for internal work, customer projects, or both?
- What would you need to trust and deliver the final output to a customer?
You can explore the application without running a project. Executing a project currently requires an Anthropic or OpenAI API key and an invitation code from me.
Repository: RamboxRoot/AgenticCompany
[link] [comments]