Cold Start¶

What happens when an agent has no history? AgentSpec gracefully handles brand new agents through manifest seeding.

The problem¶

A fresh .agent manifest has:

Declared skills (but no demonstrated proficiency)
Persona and traits (but no track record)
Model preferences (but no proven cost/quality data)

If the HR Agent (in caloron-noether) reads this profile, what does it see? Nothing useful — yet.

The solution: seed from manifest¶

ProfileManager.create_profile() seeds the profile from the manifest itself:

mgr = ProfileManager("./profiles")
manifest = load_agent("brand-new.agent")
profile = mgr.create_profile(manifest)

The new profile contains:

1. Declared skills (low confidence)¶

Each manifest.skills entry becomes a SkillProof:

SkillProof(
    skill="web-search",
    level="declared",          # not yet demonstrated
    evidence_type="manifest",
    evidence="Declared in .agent manifest v1.0.0",
    confidence=0.3,            # low — not yet proven in real sprints
)

2. Bootstrap memory (signed)¶

A signed memory captures the agent's identity:

content: "Agent initialized from manifest. Persona: precise-researcher.
          Traits: cite-everything, never-guess. Capability: reasoning-high."
category: caloron:evolution.change
status: validated
confidence: 1.0

3. Model preference memory (signed)¶

For HR Agent context:

content: "Preferred models: claude/claude-sonnet-4-6, gemini/gemini-2.5-pro,
          local/llama3:70b. Fallback: reasoning-mid."
category: caloron:professional.tool_knowledge
status: validated
confidence: 1.0

After the first sprint¶

When skills get used in a real sprint:

mgr.process_retro(profile, feedback={
    "tools": ["web-search", "cite-sources"],   # actually used
    "assessment": "completed",
}, sprint_id="sprint-1")

The matching skill proofs upgrade:

Before:  web-search: declared (30%)
After:   web-search: demonstrated (70%)

Skills not used stay at declared (30%) until their first real sprint.

What the HR Agent sees¶

Cold start (sprint 1):

Agent: deep-researcher
  Sprints: 0
  Skills (3, all declared at 30%):
    web-search, cite-sources, summarize
  Memories: 2 (bootstrap, model preferences)
  Portfolio: empty

After sprint 1:

Agent: deep-researcher
  Sprints: 1
  Skills (3, 2 demonstrated):
    web-search: demonstrated (70%)
    cite-sources: demonstrated (70%)
    summarize: declared (30%)        # not used yet
  Memories: 4 (bootstrap + 2 retro findings + 1 note)
  Portfolio: 1 entry — Quantum Research, 1/1 tasks
  Completion rate: 100%

After 10 sprints:

Agent: deep-researcher
  Sprints: 10
  Skills (8, 6 expert):
    web-search: expert (95%)
    cite-sources: expert (92%)
    arxiv-mcp: proficient (85%)
    ...
  Memories: 47 (validated, signed)
  Portfolio: 10 entries across 4 projects
  Completion rate: 90%

Why this matters¶

The cold start design ensures:

The HR Agent always has something to work with — even on day 1, it knows what skills the agent declares and what runtime to use
No special-case code — the same process_retro() flow works for sprint 1 and sprint 100
Honest signaling — declared(30%) vs demonstrated(70%) makes confidence visible
Fast bootstrap — agents become useful immediately, not after some warmup period

Resetting an agent¶

To reset an agent's profile (start over):

import shutil
shutil.rmtree("./profiles")  # nuclear option

# Or per-agent
import os
os.remove(f"./profiles/{agent_id}.profile.json")

# Recreate
profile = mgr.create_profile(manifest)

The agent_hash in the profile ties it to a specific manifest version. If the manifest changes, you can keep the old profile or start fresh.