AI ticket investigation · 2026
ATR
Follow a support ticket from code, data, and logs to reviewed findings. Explore how agents and engineers work from the same saved investigation.
Role and context
- Contribution
- Investigation workflows, team workspaces, saved state, and controlled code changes.
- Setting
- Internal engineering system at PeopleStrong · 2026
- Scope
- Architecture overview only. Customer tickets, repositories, and operational records are not shared.
The problem
Support investigations need information from tickets, code, databases, logs, and product documentation. Teams have different tools and access, so the investigation has to account for what each team can use.
A closer look
Handling large repository graphs
- Constraint
- Combining every repository graph into one payload hit memory and string-size limits.
- Decision
- Each repository graph is stored separately and combined through a bounded streaming merge.
- Tradeoff
- Storage and merging became more complex, but the full graph no longer had to fit in one in-memory string.
- Not included
- This describes the storage design. Before-and-after resource benchmarks are not included.
Explore the implementation
The investigation flow covers ticket analysis, reproduction, evidence checks, review, and controlled code changes.
Team workspaces include the available tools, models, data sources, and permissions.
Automated runs and the engineering console share saved investigation state, so engineers can take over without starting again.
Code changes stay separate from diagnosis, with isolated workspaces, diff reviews, and team permission checks.
Repository graphs use separate storage and a streaming merge after large payloads hit memory and string-size limits.
Architecture overview
View architecture
- Issue intake
A support ticket starts the investigation with the information it is allowed to use.
- Team context
Team and product documentation guide the investigation. Internal instructions are not shown here.
- Evidence adapters
Read-only tools gather relevant code, data, and logs. Internal integrations are not shown here.
- Investigation
Agents check possible causes against evidence and save their findings.
- Review gate
A separate review checks the conclusion and confidence against the evidence.
- Human handoff
An engineer can continue the saved investigation with its evidence.
- Guarded action
Code changes require separate permissions, an isolated workspace, and review.
Select a step to explore the flow.
Design decisions
Team-specific access
Each investigation uses only the data sources, tools, and permissions available to its team.
Saved investigations
Evidence, delegated work, and conversation state are saved for recovery and human review.
Diagnosis review
A separate review step checks the evidence before a diagnosis is finalized.
Controlled changes
Code changes require an isolated workspace, review, and permission checks.
Results & limits
Scope and results
The system supports saved evidence, human review, team-specific access, and separately controlled code changes. Diagnosis quality needs to be measured against reviewed ticket outcomes. Outcome rates are not published here.
Takeaway
Saving the full investigation state made recovery and engineer handovers easier.