Context Graphs Enable Agents and Humans to Act Intelligently
Summary
At the end of last year, Jaya Gupta and Ashu Garg from Foundation Capital published a great topical post that delved into the potential of context graphs. The post, titled “AI’s trillion dollar opportunity: Context Graphs” was placed in direct dialogue with Jamin Ball’s piece, “Long Live Systems of Record,” which effectively advocated for the continued primary role for systems of record in enterprise tech stacks. Both posts have made the rounds and are worth a read – especially if you are an enterprise leader and are actively wrestling with the tension between existing systems and processes and a rapidly approaching agent-centric future.
In sum: Ball posits that ‘systems of record’ will evolve, and be unbundled and rewired into warehouse, semantic, and governance-centric ‘truth layers’ that provide direction to agents through clear rules. In enterprise organizations, they will continue to function as the ‘source of truth,’ where canonical data lives and can be accessed, increasingly by fleets of agents built to execute transactions. Ball does grant that data warehouses become the ‘natural substrate’ for organizations to ‘encode the precedence and meaning’ of data, including the reconciliation of conflicting definitions or rules. In our assessment, the problem he identifies is the right one – but we would offer up a slightly different solution.
Gupta and Garg point to a different challenge – if we want agents to reason (versus simply execute tasks and rote transactions), the real missing layer is not better access to data warehouses or governance. Instead, agents need durable, queryable records of how decisions were actually made - including exceptions, approvals, and precedent. They propose that agent orchestration layers will become the new systems of record by capturing these “decision traces” into a context graph, ultimately forming the most defensible source of truth for enterprise autonomy.
At Parable we’ve been deeply contemplating - and building - context graphs for the last year and a half, and can provide some perspective based on our real-world experience. The Foundation Capital write-up is compelling but our conviction in the value of context extends even further.
We agree systems of record and governed data warehouses won’t be sufficient for underpinning true enterprise intelligence; instead, we are pursuing a system of understanding.
Gupta asserts that “Agents don’t just need rules, they need access to the decision traces that show how rules were applied in the past, where exceptions were granted, how conflicts were resolved, who approved what, and which precedents actually govern reality.” Decision traces are one class of context, but our context graph is even broader – it models how work happens far beyond executional process transactions which I describe as “point decisions.”
Put simply, our context graph can zoom all the way out to the goals and and the proportional investments in broader themes of work which define operational reality or zoom all the way down into the relationships between atomic units of event data that coalesce around decision time. This is because our graph was borne out of the problem statement we are trying to solve: how can we contextualize work across all siloed data sources, so that we can create an observability platform for operational reality? Parable measures the attention economy of an enterprise and presents those findings back to leadership. If there are interventions in the operating model (agents, change management etc.) Parable measures if they indeed produced business value.
If agents understand the decision traces (including the exceptions), the dynamic stream of activities and outputs, plus how those traces fit into a bigger picture of organizational context, they will simply be more viable and much more effective.
Systems of record are inherently limited as we move to a ‘declarative’ paradigm
Systems of record are deeply entrenched in organizations, despite largely failing at their most fundamental job: to be the source of truth. There are technical constraints (they are reliant on user instruction and step-by-step commands, both in the UI and in the database design) as well as constraints imposed by the reality of the workplace: work as we know it is collaborative and horizontal, and systems of record are vertical.
As we move to a ‘declarative’ paradigm – where users describe the desired outcome and let the system (or agent) figure out the necessary steps - the context for why a job is being completed beyond the transaction itself (and even beyond variants of the transaction history) is paramount to the agent’s efficacy. In our view, data warehouses do not solve this problem because they do not understand relational context in a structured way, and they do not perform cross-system synthesis. Humans or agents executing joins in SQL tables cannot address this. It needs to be architected bottoms up.
Effective agent + human collaboration requires context, appreciates nuance, and produces explainability
Agents do need rules, but rules are low-dimensional while the reality of operating in an enterprise is high-dimensional. Our view is that for agents to be successful they not only need decision traces - they need broader contextual material to honor rules and goals against a given backdrop. Humans-in-the-loop will need explainability, too, to interrogate why an agent made a given decision and course-correct if needed. That human worker may need to codify specific data definitions for a clear and obvious prompt in one scenario, or apply ad-hoc definitions for a different exploratory or situational analysis. This is made possible by a data architecture designed with this workflow in mind.
Better governance, semantic contracts, and explicit rules for which definition prevails is unrealistic for all but the most obvious use cases. No organization has ever comprehensively identified every case and successfully written all the rules for data governance. Context is what reconciles inputs at decision time, and creates leverage for the person (or agent) responsible for making said decision. Put simply, agents will never be intelligent actors in a perfectly structured environment - they operate in a busy bazaar of context, counterfactuals, relationships, and dynamically evolving inputs. Which means decisions - and their consequences - live beyond the immediate isolated transaction. Humans solve this by processing broader organizational context as well as transactional context to make point decisions. CS teams manage trade offs between customer happiness and cost and sales teams manage effort trade offs between enterprise and mid market prospects in the pipeline. This context is critical for understanding the variances we find in process flow and the context for these decisions lives outside of process streams.
Parable’s context graph assembles a current picture of reality in the workplace
A final point here: Arvind Jain from Glean astutely observed in their posts that enterprises “can’t reliably capture the why; you can capture the how.” Our perspective is that there is a lot of rich insight produced between those two poles, where the “how” = process and the “why” = intention. Parable effectively captures the “how” and the “how much” + “where” + “what” + “who” + “when” – and in doing so, builds a context graph which can effectively provide organizations with clear observability of operational reality and where it’s frictional.
Interestingly, when we provide enterprise leaders with this X-Ray vision, they can imbue that analysis with the “why” almost instantly. This doesn’t solve every problem in building effective agents - but it does two things very well:
- It programmatically identifies and quantifies operational inefficiency sharpening how agents should be deployed.
- It precisely measures the business impact of efficacy of every agent deployed against a pre-established baseline to ensure it’s producing genuine value.
Moving from here to there, and going back to first principles
We’ve spent a lot of time discussing context graphs and agents, when the reality is that so many enterprises are fairly early in their AI transformation journey. Context graphs are critical to an agentic agenda, but we can also use ours to effectively autopsy any business today – and create significant operating leverage for leaders as a result.
Our context graph can answer the questions on every leader’s mind: how does work happen now, as a baseline and where are we wasting time? Where and how much should we invest in AI to improve this baseline? Did those AI investments drive meaningful and measurable business value? How do we redeploy that value into growth initiatives?
Parable is not trying to replace systems of record; it becomes the system of understanding that sits between fragmented data truths and real context, giving humans the data they need to leverage agents intelligently.
