FAIR Contexts Maturity Model
The key to managing agents is managing context
The FAIR Contexts Maturity Model (FCMM) extends the FAIR Guiding Principles from Scientific Data — Findable, Accessible, Interoperable, Reusable — to the context flows that AI agents read, write, and use. The key claim is that “acceptable data quality” can only be judged at the point of use, which requires each agentic workflow to take responsibility for evaluating its input, outputting its provenance, and feeding learnings back into steering.
Why we need this
Agentic workflows are more like biological systems than deterministic software. Their behavior is unpredictably sensitive to their context, not just their inputs. Extending their usage to new contexts requires the equivalent of “post-market surveillance” (PMS, aka Phase IV) to continually identify and remediate pathological outcomes before they spread.
By taking responsibility for managing contexts as first-class objects, like code and data, organizations can accelerate AI experimentation and adoption while still bounding negative outcomes.
FAIR Contexts
FAIR Contexts extend the four FAIR principles to keep up with the velocity and needs of modern, AI-driven workflows.
F — Findable
- Searchable via metadata, semantics, and full-text
- Linkable by persistent, versioned identifiers
A — Accessible
- Any entity can read and write all and only the contexts it should
- Uniform affordances for reading, writing, and querying structured and unstructured data
- Artifacts are directly legible to both humans and machines
I — Interoperable
- Not tied to any particular tool, vendor, or platform
- Works seamlessly with any LLM, harness, or cloud provider
- Can be efficiently shared, even across organizations and platforms
R — Reusable
- Carries “provenance on write” so consumers can make their own trust decisions
- Includes link-backs to source inputs (“Chain of Context”)
- Can function as Outputs, Inputs, Steering, or Learnings (“OISL”)
Initially, you can check artifacts using existing FAIR tools (e.g., FAIRshake). Over time we expect the community to develop more agent-native tooling.
The maturity ladder
Because of the contextual nature of agentic work, the atomic unit of maturity is the “flow.” A team, process, or organization is only as mature as the weakest flow that feeds into its business outcomes.
FCMM-0 — Awareness
Do we treat context as a valuable, persistent, compounding asset?
FCMM-1 — Stewarding
Does each flow have an Agent Context Steward, who is responsible for:
- ensuring the FAIRness of outputs, inputs, and steering
- configuring and monitoring the eval/learning loop
- communicating with upstream and downstream stewards
FCMM-2 — Provenance
Does every Output from this flow carry its full versioned provenance, e.g.:
- inputs
- steering
- agent harness
- model family
- tool use
FCMM-3 — Monitoring
Are runtime errors emitted, tracked, monitored, and alerted on?
FCMM-4 — Learning
How well does the system self-correct?
- Are outputs independently and continually judged against inputs to detect context drift?
- How often do bad outputs invisibly poison downstream flows?
- How efficiently are learnings fed back into the steering?
- How quickly can humans discover when quality has degraded?