Skip to content

README.md

Latest commit

 

History

History
136 lines (108 loc) · 5.89 KB

README.md

File metadata and controls

136 lines (108 loc) · 5.89 KB

Agent Overview

The TelemetryFlow Hermes multi-agent team uses Hermes profiles — each agent has its own SOUL.md, memory, skills, config, and Telegram bot. They share nothing.

Team Architecture

graph TD
    ALERT["Alert Fired<br/>TFO Webhook / Telegram"]
    TRIAGE["<b>Triage</b><br/>Paranoid Gatekeeper"]
    INV["<b>Investigator</b><br/>Hostile Scientist"]
    REV["<b>Reviewer</b><br/>Professional Skeptic"]
    REM["<b>Remediator</b><br/>Cautious Pragmatist"]
    HUMAN["<b>Human</b><br/>Approve / Reject"]

    ALERT --> TRIAGE
    TRIAGE -->|"genuine anomaly"| INV
    TRIAGE -->|"INCOMPLETE"| CHALLENGE["Issue challenge<br/>to Investigator"]
    TRIAGE -->|"known pattern"| RESOLVED["Auto-resolve<br/>(from MEMORY.md)"]
    TRIAGE -->|"noise"| DISCARD["Discard"]
    CHALLENGE --> INV
    INV --> REV
    REV -->|"approved"| REM
    REV -->|"rejected"| INV
    REM -->|"requires_approval"| HUMAN
    HUMAN -->|"approve"| EXECUTE["Execute Action"]
    HUMAN -->|"reject"| ESCALATE["Escalate"]

    style TRIAGE fill:#1a6b4a,stroke:#00d4aa,color:#fff
    style INV fill:#1a3a6b,stroke:#3b82f6,color:#fff
    style REV fill:#6b4a1a,stroke:#f0883e,color:#fff
    style REM fill:#6b1a2a,stroke:#ef4444,color:#fff
Loading

Agent Profiles

All agents operate as adversarial scientists — they assume inputs are wrong, actively falsify hypotheses, and never trust without proof.

Agent Model Max Turns Role Tools
Triage glm-5.1 (OpenCode Go) 30 Paranoid gatekeeper terminal, web, delegation
Investigator claude-sonnet-4-5 (Anthropic) 45 Hostile scientist terminal, web, delegation
Reviewer glm-5.1 (OpenCode Go) 20 Professional skeptic terminal, web (read-only)
Remediator glm-5.1 (OpenCode Go) 15 Cautious pragmatist terminal, web (approval required)

Delegation Model

sequenceDiagram
    participant T as Triage
    participant I as Investigator
    participant R as Reviewer
    participant M as Remediator
    participant H as Human

    Note over T: Alert received
    T->>T: Read MEMORY.md
    T->>T: Classify: critical/known/noise/INCOMPLETE
    alt Known pattern
        T->>T: Auto-resolve
    else INCOMPLETE classification
        T->>I: delegate(task, context, challenge)
    else Genuine anomaly
        T->>I: delegate(task, context)
        Note over I: Fresh context, full tool access
        I->>I: query_metrics + search_logs + list_traces + get_exemplars
        I->>I: Falsify hypotheses — guilty until proven innocent
        I->>I: Document dead hypotheses
        I->>R: delegate(evidence, surviving hypothesis)
        Note over R: Separate context, read-only tools
        R->>R: Hunt for disconfirming evidence
        R->>R: Verdict: CONFIRMED / NEEDS_MORE_EVIDENCE / REJECTED
        R->>M: CONFIRMED hypothesis + recommended actions
        Note over M: Three gates: root cause, proportional, approval
        M->>M: Blast radius analysis
        M->>H: Telegram notification
        H->>M: Approve
        M->>M: Execute remediation
        M->>M: Post-action verification
    end
Loading

Profile Configuration

Each profile lives at profiles/<agent-name>/ with:

File Purpose
config.yaml Model, turn cap, tool permissions, gateway config
SOUL.md Agent identity and behavioral constraints
memories/MEMORY.md Persistent agent facts (2,200 char max)
memories/USER.md User profile and preferences (1,375 char max)

Profile Isolation

  • Separate Telegram bots — each agent has its own TELEGRAM_BOT_TOKEN_*
  • Separate memory — agents don't share MEMORY.md or state.db
  • Separate skills — each profile can have different skill sets
  • Separate config — different models, turn caps, tool permissions

Context Management

Why Separate Contexts?

A single agent doing everything fills its context window fast. By step 4 it's operating on summaries of summaries, and quality degrades. Separate concerns: each agent has one job, one context, one set of tools.

Reviewer Bias Prevention

The Reviewer runs in a completely separate context with:

  • Zero visibility into the Investigator's thought process
  • Read-only tools only (can't modify investigation)
  • Fresh perspective on evidence and hypothesis

This prevents:

  • Confirmation bias — defending the Investigator's conclusions
  • Anchoring bias — overweighting initial findings
  • Sunk cost bias — continuing a failed investigation approach

Cost Model

Agent Model Turns/Incident Cost/Turn Cost/Incident
Triage glm-5.1 3-5 ~$0.002 ~$0.01
Investigator claude-sonnet-4-5 10-20 ~$0.008 ~$0.05-0.15
Reviewer glm-5.1 5-8 ~$0.005 ~$0.03-0.08
Remediator glm-5.1 2-5 ~$0.003 ~$0.01-0.03
Total 20-38 ~$0.10-0.27

Detailed Agent Docs