Download

AI and agent builders

Test the context before your agent acts on it.

Your agent is only as good as what you tell it. Subjourney runs your instructions, knowledge base and process docs as journeys, with simulated users and the agent's own path, so you find the gaps in context before they become wrong answers in production.

Run your agent's context
  • AI engineers writing system prompts and tool instructions that have to hold up in the wild.
  • Context and knowledge owners whose docs agents now retrieve and act on.
  • Product teams shipping agents who need to know what the agent will do with the user's real goals.
  • Teams building with coding agents who want their specs and context files tested before the agent builds from them.

Agents don't ask clarifying questions. They guess.

A person who hits a gap in a doc stops and asks. An agent fills it, confidently, and moves on. Every ambiguous instruction becomes a behaviour you didn't choose.

Evals tell you the output was wrong. They rarely tell you which line of context made it wrong, or what was missing.

What a run catches

  • Missing

    The unstated rule.

    Your support agent's instructions cover refunds and cancellations, but not the customer who wants both.

  • Missing

    The knowledge gap.

    The agent is told to “follow the escalation policy”. The escalation policy isn't in its context.

  • Contradictory

    The instruction conflict.

    The system prompt says “never promise timelines”; the help-centre article it retrieves promises 48 hours.

  • Dead end

    The goal it can't reach.

    A user persona with a reasonable request and no path through the agent's tools and docs to get there.

Everything your agent reads before it acts.

  • System prompts and agent instructions.
  • Knowledge-base and help-centre articles the agent retrieves.
  • Tool and API descriptions that shape what the agent can do.
  • Process docs and policies the agent is meant to follow.
  • Context files for coding agents: specs, rules files, project docs.

Evals tell you it failed. A run tells you where the context broke.

  • A wrong answer

    Evals alone
    You know the output failed
    Evals plus context runs
    You see the instruction or doc that caused it

  • Missing context

    Evals alone
    Shows up as a hallucination in production
    Evals plus context runs
    Shows up as a stall, traced to the gap

  • Conflicting sources

    Evals alone
    Agent picks one, silently
    Evals plus context runs
    Conflict flagged, both paragraphs named

  • Test cases

    Evals alone
    Written one at a time
    Evals plus context runs
    Personas generated from the journeys your context describes

  • Context changes

    Evals alone
    Re-run the eval set and diff the scores
    Evals plus context runs
    Re-run the journeys and see exactly what broke

Run your context the way you run your tests.

  1. 1Import the system prompt, the knowledge base and the processes the agent supports.
  2. 2Run user personas with real goals through the journey the agent is meant to serve.
  3. 3See each stall traced to the instruction or document responsible.
  4. 4Send findings to your coding agent to patch the context.
  5. 5Re-run on every change, before it ships.

In practice

A team building an onboarding agent runs its context against five new-customer personas. Four reach a working setup. The fifth, on a legacy plan, stalls at data import: the agent's docs only describe the current plan's limits. The finding goes straight to their coding agent, the docs get a legacy-plan section, and the re-run passes.

A run, step by step

One agent, five users, the gap evals didn't name.

  1. 01Import.

    The team brings the onboarding agent's system prompt, twelve help-centre articles it retrieves, and the plans-and-limits doc.

  2. 02Model.

    Subjourney builds the journey.

    Get started
    Create workspace
    Connect data
    Import dataFlagged
    Invite team
    First result

    It flags that Import data depends on plan limits that only one article describes.

  3. 03Run.

    Five simulated personas set off. Choose one to see what they want and what gets in their way.

    CSV userAnalyst

    A CSV, not a live connection.

    Motivation
    Try it with the spreadsheet already in hand, not a live database.
    Frustration
    Import steps that assume a connection IT hasn't approved.
    Need
    A CSV upload that's offered, not hidden behind connectors.
    • Age44
    • Income$78,000
    • EducationBachelors Degree
    • EmploymentMid Career
    • Communication styleFormal
    • Social engagementPassive Observer
    • Risk appetiteRisk-Averse
    • Eco awarenessPassive Concern
    • Work-life balance9 to 5
    • Brand loyaltyHabitual
    • Price sensitivityCost-Conscious
    • Tech adoptionLate Adopter
    • Tech literacyIntermediate

  4. 04Findings.

    ContradictoryStalled at: Invite team

    Admin asking for SSO

    The system prompt says to route SSO questions to sales. The SSO help article tells users to enable it themselves in settings.

    Open in sourceSend to agent
    MissingStalled at: Import data

    Customer on a legacy plan

    The plans doc covers current plans only. The agent has no basis for legacy import limits, so it would guess.

    Open in sourceSend to agent

  5. 05Fix and re-run.

    Findings go to the team's coding agent, which adds a legacy-plan section and aligns the SSO guidance. Re-run: all five personas reach a first result.

    The eval would have caught a wrong answer eventually. The run named the paragraph first.

Personas to start with

Run the users your agent will actually meet.

  • The clear request.Knows what they want and asks for it directly.
  • The vague request.Describes a problem, not a task. Does your context help the agent work it out?
  • The two-part request.Wants something your instructions only cover separately.
  • The out-of-scope request.Should be declined or handed off. Does the context say how?
  • The persistent user.Pushes back after the first answer. Does the agent's guidance hold?

Questions

Does this replace our evals?

No. Evals score outputs. Subjourney tests the context behind them, so when an eval fails you know which line to change.

What can I import?

System prompts, knowledge-base docs, process docs and specs: anything your agent reads or acts on.

Can findings go straight to my coding agent?

Yes. Send findings to the agent you build with, then re-run to verify the fix.

Your agent's context is code. Test it like code.

Download Subjourney