AI Research / Agentic operating systems
Design how agents work together.
An agentic operating system brings priorities, context, authority, coordination, and review into a working whole. Our research examines those operating conditions and how they affect the work a team can actually accept.
Discuss an operations studyParallel work converges through review
Agree the intended result, boundaries, dependencies, and the evidence the receiver needs.
Read every step
- Set a shared goal. Agree the intended result, boundaries, dependencies, and the evidence the receiver needs.
- Assign bounded work. Participants receive distinct assignments and sufficient context. Dependencies determine what can proceed.
- Integrate and inspect. Bring the packets together, resolve overlaps, and return incomplete work before acceptance.
- Deliver with evidence. The receiving owner checks the integrated result and knows what remains unresolved and who owns the next step.
The work system is the subject of the research.
Here, “operating system” describes how human and agent work is organized. It includes planning, the transfer of work between participants, and the records needed to inspect and recover that work. Parallax is an environment for developing and examining these practices.
This is an applied research agenda. Internal implementation and a successful demonstration do not establish effectiveness across customer environments.
Planning that connects priorities to ready work.
Scrum offers useful questions about goals, backlogs, review, and adaptation. Agent work adds another question: when a dependency clears, what is ready to move now?
We investigate how a human review cadence can coexist with smaller units of agent delivery. Teams may plan around a shared goal while releasing work when its inputs and authority are ready, rather than making every participant wait for a batch.
Study question: Which planning and review cadence reduces waiting without creating unstable priorities or excessive coordination?
Measures: time waiting, time to accepted work, priority changes, review effort, and unfinished work. Comparison requires an agreed baseline; faster dispatch alone is not evidence of a better result.
Work packets that a receiver can use.
A packet transfer is a handoff of a bounded unit of work. The recipient needs enough context to proceed and a clear way to return an incomplete or stale assignment.
Illustrative work-packet contents
- The intended result and acceptance criteria.
- The relevant sources, versions, and unresolved questions.
- The scope, dependencies, permitted actions, and accountable owner.
- The expected return: work produced, evidence, limitations, and next decision.
- A recovery path when inputs are missing or authority has changed.
This is a conceptual checklist, not an exported internal protocol or a claim that every runtime enforces it.
Study question: What is the smallest sufficient handoff for a receiver to continue correctly?
Measures: context volume, missing-input refusals, repeated discovery, stale-source errors, and handoff acceptance. The objective is sufficient context with less reconstruction.
Orchestration that improves the delivered result.
Multiple agents can investigate, implement, and review different parts of a problem. They also create dependencies, overlapping edits, duplicated effort, and integration work.
Our agenda examines bounded assignments, shared contracts, work ownership, review, and recovery when a participant fails. Additional agents are useful only when their contribution exceeds the coordination they introduce.
Study question: When does a coordinated group outperform a smaller team on the same accepted scope?
Measures: accepted output quality, elapsed time, total operating cost, rework, collisions, and recovery effort. Compare equivalent tasks and report the workload and limits alongside the result.
Bring a workflow we can examine together.
A scoped study could compare two planning approaches, test a handoff format, or evaluate a swarm against a simpler delivery approach. Begin with a permitted task set, an owner who can judge the result, and a budget for the investigation.
The proposal defines the study brief, experiment records, result review, and recommendations. Production integration, continued service access, and technology rights are agreed separately.
Compare research funding and collaboration options
Explore Parallax · Explore Decision Thread · Explore Agent Centered Design
Tell us what needs to work better.
Discuss an operations study