Back to the blog
HELIX / ENGINEERING GRAPHS

Engineering graphs for agentic software

How graph-shaped context gives coding agents better retrieval, safer plans, and feedback loops that improve with every engineering event.

Engineering GraphsAgentic Systems

Coding agents are usually given a repository and a prompt. That is enough to generate code. It is not enough to understand a software organization.

The missing information lives in relationships: the service that owns a file, the incident that led to a guardrail, the pull request that introduced a dependency, the test that supports a requirement, and the team that will absorb the operational risk.

Context is connected

Vector search is useful for finding text that resembles a question. Engineering questions often require a different operation: following a path.

Consider “What could changing the checkout client break?” A useful investigation may need to traverse:

text
checkout-client.ts
  → IMPORTED_BY → Checkout API
  → CALLS → Payment Gateway
  → OWNED_BY → Commerce Platform
  → COVERED_BY → payment-retry.spec.ts
  → CHANGED_IN → PR #418

No single document contains that answer. The answer emerges from the path.

Model the software and the work

An Engineering Graph should represent more than static code structure. Its nodes can include:

  • repositories, services, modules, APIs, files, and dependencies;
  • pull requests, commits, releases, incidents, and architecture decisions;
  • people, teams, ownership boundaries, and review history;
  • agent runs, requirements, tests, findings, and production outcomes.

Relationships give those nodes meaning. IMPORTS, DEPENDS_ON, AUTHORED, LED_TO, VERIFIES, and AFFECTED_BY answer different questions and must not be collapsed into generic “related” edges.

Let agents operate on subgraphs

The full organizational graph is too large and noisy for every task. Helix builds an explainable subgraph around the user’s intent:

  1. resolve the entities named or implied by the question;
  2. choose relationship types relevant to that intent;
  3. traverse a bounded number of hops;
  4. rank paths by directness, evidence quality, and recency;
  5. give the language model structured facts, not an undifferentiated text dump.

The model can then synthesize a conversational answer while the graph preserves provenance.

Close the learning loop

The graph should not be a snapshot. It should continuously learn from engineering work.

EventGraph updateFuture benefit
Pull request mergedConnect changes, owners, findings, and reviewersBetter change-risk context
Incident resolvedConnect symptoms, causes, fixes, and servicesFaster diagnosis
Agent run verifiedConnect actions, requirements, and test evidenceSafer autonomous work
Engineer corrects a claimPreserve correction and supporting sourceBetter future answers

Every answer can also reveal a knowledge gap. If Helix cannot establish ownership or find evidence for a critical dependency, that uncertainty becomes a first-class graph record. The organization now knows what context it is missing.

Build for humans and agents together

Humans need a view that makes complex systems understandable. Agents need a structured context layer they can query before acting. Both need claims that can be traced back to evidence.

That shared requirement is why the Engineering Graph is the core of Helix. It is not a visualization added to a chatbot. It is the living system of record that makes useful questions, safer plans, and independent verification possible.