As multi-agent systems become more capable, I’m starting to wonder whether we need a different approach to project management.
Most project management tools are built around humans as the primary workers. But with agents handling research, coding, testing, deployment, and other tasks, the platform may need to manage agents as well.
A useful project management platform for AI agents could provide:
Tasks assigned to specific agents
Shared project context and memory
Dependencies between agent tasks
Parallel execution with clear ownership
Logs of agent actions and outputs
Human review and approval where needed
The recent HF Agent OS project made me think about this from another angle: if agents can already manage parts of the Hugging Face project lifecycle, perhaps the next layer is managing the actual work they perform.
I’ve been looking at Sharkly.ai, which approaches this by putting AI agents and human teammates into the same project workflow.
For those building with Hugging Face, smolagents, or other agent frameworks: what would you expect from a project management platform designed specifically for AI agents?
HF Agent OS actually already does a slice of this, Planner classifies intent and routes it, Orchestrator dispatches to specialized sub-agents and logs task lifecycle, that’s project-management structure baked directly into an agent framework rather than bolted on top. Worth reading through that thread since it’s a working example of what you’re describing, not just theory.
On what a PM layer for agents actually needs beyond what humans need: dependency tracking has to be much stricter since agents don’t self-correct on ambiguous ownership the way a human teammate would just ask a clarifying question, so “task assigned to agent X, blocks task Y” needs to be enforced structurally, not just documented. Logs/traceability matter more too, since debugging why an agent made a decision requires the full context it had at that moment, not just a task status update. One thing to flag since it came up in your other post as well, if you’re specifically evaluating Sharkly.ai for this, worth being upfront that you’re asking about a specific product rather than framing it as open research, that’ll get you more useful comparisons from people who’ve tried similar tools.
A project-management layer becomes useful when it manages claims and evidence, not only task cards. Each agent task should carry an owner, capability scope, input snapshot, dependencies, budget, stop condition, and a verifier for the expected end state. Human review can then sit at risk boundaries instead of every step. The hard cross-tool problem is durable agent identity plus an audit trail that survives handoffs and retries.