“We went from AI producing plausible answers to evidence-backed work.”
3Victors built VictorBot, an in-house agent that started as an AWS error dump in Slack and now investigates production issues, validates Jira stories, and opens reviewed PRs. Unblocked is the context layer underneath it, supplying the institutional knowledge from Slack decisions, PR history, Jira, and Confluence that turns a detection tool into one the team trusts to work on its own. It runs roughly 8,000 Unblocked calls a month.
8,000
Unblocked calls a month behind one internal agent
~90%
of work is diagnosis and validation, not new code
0
tokens spent on routine monitoring checks
3Victors, an ATPCO company, is a first-of-its-kind travel data analytics company, transforming how the travel industry harnesses the power of data and leverages AI. It runs a data platform on AWS that its engineers monitor and extend every day.
Like a lot of teams in 2026, they wanted agents doing real work. Unlike most, they built the agent themselves.
VictorBot started narrow. It was a monitoring dashboard that pulled scattered AWS signals into one place: SQS backlogs, ECS task failures, Step Functions failures, and CloudWatch error bursts. "The initial goal was to make these alerts easier to see and act on," says Farzwan Mohamed, a software developer at 3Victors who works across application development, cloud infrastructure, data workflows, and production monitoring. Built on Hermes as an open source agent runtime, with connectors layered on top, it has since become something else entirely.
The challenge: an agent you couldn't leave alone#
When Farzwan joined in October 2025, the ceiling was obvious.
"The blocker was giving AI enough accurate context to produce useful work without hallucinating, missing important constraints, or generating low-quality output," he said. "It was difficult to trust an agent to work independently for an extended period and expect production-ready results."
That gap capped what VictorBot could be. Detection was solved. Anything past detection required knowing why the system was built the way it was, what had already been tried, and which past decisions still applied. None of that lives in the code or the CloudWatch logs.
"Once we had connected AWS to an agent with Unblocked access, the obvious next question was: why stop at detection?" Farzwan said. "The action layer was built around this."
The solution: an evidence chain, not a prompt#
Today, a typical VictorBot run starts with a Jira story number and ends with a reviewed PR. Eight steps, and context comes first.
- Pull the ticket through the Jira MCP: description, acceptance criteria, comments, linked stories, and relevant Confluence pages.
- Pull the institutional context through Unblocked: whether the feature has been discussed in Slack, attempted in another PR, or constrained by an earlier design decision.
- Inspect the current implementation across the relevant GitHub repositories: handlers, models, configuration, tests, deployment manifests, and cross-repo dependencies.
- Check the running system where it matters. If the story concerns runtime behavior, VictorBot checks live AWS state or queries the database. As Farzwan puts it: "Code explains intended behavior. Live state proves what is actually happening."
- Build an acceptance-criteria matrix. Each criterion is classified as covered, partially covered, missing, ambiguous, or not a code concern.
- Choose an outcome. Sometimes it's a test plan or a confirmed configuration issue. If code is required, it creates a feature branch, makes the smallest scoped change, and runs the relevant tests.
- Review the change with an independent reviewer agent that checks security, logic, schema compatibility, deployment consequences, and acceptance-criteria coverage.
- Open the PR, explaining the ticket, the implementation, the tests, the deployment impact, and any unresolved decisions.
Step two, where Unblocked shines, is the point of leverage for VictorBot.
"Unblocked gives VictorBot the why, who, and when," Farzwan said. It retrieves from Slack discussions, GitHub code and PR history, Jira issues and comments, Confluence documentation, and past incidents and design decisions to answer the questions code search can't:
- Was this behavior intentional?
- Has a rollout like this caused previous incidents?
- Has another team already implemented something similar?
- Was an earlier approach rejected?
- Which PR introduced this convention?
- Is the Jira story missing context that exists in Slack?
"Instead of treating AI as a tool for isolated prompts or code generation, we now give it access to the full evidence chain: live system state, code, operational history, documentation, and verification steps. It's possible to investigate an issue, understand its context, implement a change, test it, and validate the result within a single conversation."
Farzwan Mohamed — Software Developer, 3Victors
Most of the work isn't writing code#
Roughly 90% of the time, VictorBot doesn't change a line.
"A large share of the work is diagnosis and validation rather than writing new code," Farzwan said. The agent gathers evidence from the running system, the existing implementation, the documentation, and prior decisions to confirm whether a reported problem is actually a defect and where it originates.
When code is needed, it looks for established patterns and reusable implementations in the existing codebase before writing anything new. "Because the work is grounded in existing code and validated before changes are made, the generated code rarely needs significant revision."
It only fixes something when it has a clear requirement, a reproducible mismatch, or explicit permission. Otherwise it flags for human review: when acceptance criteria are absent or ambiguous, when product intent is unclear, when several valid designs exist, when live evidence conflicts with documentation, when the issue is data or configuration rather than code, or when the change would cross a policy, security, or infrastructure boundary.
That's the difference between a useful agent that ships and an agent that guesses and wastes human review time.
The system writes its own runbooks#
After VictorBot solves a complex or recurring problem, that workflow can become a skill on its own. On the next similar task, Hermes loads the skill before acting. If a step turns out to be stale or incomplete, VictorBot patches it.
"The procedure improves through real use rather than remaining a static runbook," Farzwan said. And it tells you when it has, at the end of most Slack sessions.
The skills are stored as markdown, so they can be reviewed and shared across teams. "A new engineer or agent no longer starts with a blank slate."
The same instinct shows up in how the team spends tokens. Most monitoring work is deterministic. Checking a queue count, an open PR, or a CloudWatch exit code doesn't need a language model. So monitoring runs in three tiers: Python scripts collect metrics and apply thresholds, and quiet runs emit nothing and consume no tokens; a state and dedup tier suppresses repeats and known benign signals; the LLM tier runs only when a new, actionable anomaly needs explaining.
Faster, cheaper, and quieter, with the model reserved for the part that actually requires judgment.
Adoption: 8,000 Unblocked calls a month#
Nobody mandated VictorBot.
"The first dashboard proved the value of pulling scattered AWS signals into one practical place," Farzwan said. "From there, adoption followed a simple loop: someone saw VictorBot surface a useful answer or shorten an investigation, then asked, 'Could it do this for my workflow too?' Each request became a new integration, reusable skill, or investigation pattern."
The scope grew well past monitoring: Figma-ready designs and visual artifacts, database investigations that need careful joins and production-safe query constraints, and AWS incident investigations across queues, ECS services, logs, and orchestration workflows. It now runs as a personalized workspace for each user in Slack.
Underneath all of it, roughly 8,000 Unblocked calls a month.
"Without Unblocked, VictorBot would not be able to resolve relevant context nearly as quickly during an investigation or development workflow," Farzwan said. "The call volume represents frequent retrieval of institutional knowledge, including prior decisions, related code, documentation, and past incidents, rather than raw usage for its own sake."
Farzwan indicated that it changed how quickly the team, not just the engineers, can understand an unfamiliar part of the codebase, validate an issue, and make business decisions with the right context.
He went on to say, "Before VictorBot, diagnosing an incident often meant jumping between Slack channels, AWS consoles, CloudWatch logs, deployment history, and database queries to piece together what was happening. That repeated shift, from a scattered, manual investigation to a grounded answer in minutes, is the clearest proof of what VictorBot with Unblocked changed."
A loop that compounds#
3Victors keeps an Obsidian vault as VictorBot's long-term memory, with separate layers for architecture and resource context, incidents and their evidence, patterns for recognizing recurring failures, runbooks, proven database and Logs Insights queries, schema and join guidance, and which errors are benign versus actionable.
"The goal was to give the agent enough organization context to prevent hallucination and give us the most deterministic answer each time," Farzwan said.
Before an investigation, VictorBot semantically searches the vault for matching services, errors, and previous incidents. After it establishes a root cause, it writes the new evidence back. Unblocked then adds the historical layer: related PRs, prior Slack discussions, Jira decisions, earlier incidents.
"The vault turns temporary investigations into durable operational knowledge."
The loop: alert, investigation, incident note, reusable pattern, skill or runbook, faster next investigation.
Bottom line for engineering leaders#
3Victors didn't replace its stack. It built on it. Hermes runs the agent, TurboVec indexes the vault, Obsidian holds some knowledge, and Unblocked supplies the institutional context none of those can produce on their own.
Next up on Farzwan's list is Jenkins-style CI/CD for autonomous code deployments to AWS from Git.
Farzwan's advice to anyone thinking about building something similar is simple: "Experiment to fit your use case."
The part worth copying is the context engine he chose to build on, Unblocked.
Get in touch with us today to see how Unblocked can be the context layer for your engineers and your agents.