agentkit

Production primitives for autonomous AI agents.

Built from 57 sessions running a live autonomous agent managing real money for 50+ days without human intervention.

Not theory. Not a tutorial. Production-tested.

pip install bitbot-primitives

Why agentkit?

Most agent frameworks focus on tool calling and prompt chaining. That's the easy part.

The hard part is what happens when your agent runs for weeks:

• How does it learn from mistakes without forgetting successes?
• How does it know when it's in unfamiliar territory?
• How does it detect when its own signals contradict each other?
• How does it avoid repeating the same error 47 times?
• How does it know when to stop?

These problems don't appear in demos. They appear at 3am on day 23 when your agent has been running autonomously and you're asleep.

agentkit solves them.

The 8 Primitives

1. LessonsMemory

Agents that learn from outcomes, not just instructions.

from agentkit import LessonsMemory

memory = LessonsMemory("lessons.json")

# Record what happened
memory.record(
    text="Buying during extreme fear with weak momentum = loss",
    pnl_impact=-12.50,
    outcome="LOSS",
    conditions={"fear_greed": 12, "momentum": -0.3}
)

# Inject top lessons into your next prompt
prompt += memory.to_prompt_block(5)

Lessons are ranked by impact * recency * frequency. Recent high-impact lessons surface first. Old lessons decay but don't disappear.

2. EpistemicHumility

Agents that know what they don't know.

from agentkit import EpistemicHumility

humility = EpistemicHumility("knowledge.json")

assessment = humility.assess({
    "market_regime": "bear",
    "volatility": "extreme",
    "signal": "weak_buy"
})

if assessment["familiarity"] == "UNKNOWN":
    # Never seen this before — reduce action size
    size *= assessment["size_modifier"]  # 0.5x

# After acting, record the experience
humility.record(conditions)

Tracks every condition combination your agent has operated in. Returns KNOWN (3+), RARE (1-2), or UNKNOWN (0) with size modifiers.

3. ContradictionDetector

Surface conflicts before your agent acts.

from agentkit import ContradictionDetector

detector = ContradictionDetector()

detector.add_rule(
    name="momentum_vs_sentiment",
    severity="HIGH",
    check=lambda ctx: ctx["momentum"] > 0 and ctx["sentiment"] < 20,
    detail=lambda ctx: f"Momentum bullish but sentiment fearful"
)

report = detector.analyze(context)
if report["level"] == "MAJOR_CONFLICT":
    # Don't act — signals disagree
    prompt += report["prompt_block"]

Scores contradictions 0-10. CONSENSUS, MINOR_CONFLICT, CONFLICT, MAJOR_CONFLICT.

4. OutcomeTracker

Link every decision to its result.

from agentkit import OutcomeTracker

tracker = OutcomeTracker("outcomes.json")

# Record the decision
decision_id = tracker.record_decision(
    action="BUY",
    reasoning="Strong signal score, low fear",
    context={"signal": 6, "fear_greed": 25}
)

# Later, record what happened
tracker.record_outcome(
    decision_id=decision_id,
    result="WIN",
    pnl=15.30,
    learned="Low F&G entries during signal alignment work"
)

# Get win rate by condition
stats = tracker.stats_by_condition("fear_greed", buckets=[0, 25, 50, 75, 100])

5. CircuitBreaker

Automatic safety limits.

from agentkit import CircuitBreaker

breaker = CircuitBreaker(
    max_consecutive_losses=3,
    max_drawdown_pct=15.0,
    cooldown_seconds=1800
)

# Check before acting
status = breaker.check(
    consecutive_losses=current_losses,
    drawdown_pct=current_drawdown
)

if status["blocked"]:
    print(f"Blocked: {status['reason']}")
    print(f"Resumes: {status['resumes_at']}")
else:
    # Safe to proceed
    execute_action()

6. Playbook

Rules that evolve from patterns.

from agentkit import Playbook

playbook = Playbook("playbook.json")

# Add a rule learned from experience
playbook.add_rule(
    name="fear_entry",
    trigger="fear_greed < 20 AND signal_score > 3",
    action="size *= 1.5",
    source="LEARNED",
    evidence={"wins": 7, "losses": 2, "avg_pnl": 8.50}
)

# Check which rules apply
applicable = playbook.evaluate(context)
for rule in applicable:
    print(f"Apply: {rule['name']} → {rule['action']}")

# Confirm or reject rules based on outcomes
playbook.confirm_rule("fear_entry", outcome="WIN", pnl=12.0)

Rules have states: CANDIDATE → CONFIRMED → PROVEN. Poor performers get demoted or removed automatically.

7. Reflexivity

Detect when momentum creates its own reality.

from agentkit import Reflexivity

reflex = Reflexivity()

# Feed it market data
score = reflex.update(
    price=67000,
    volume=1200000,
    sentiment=72,
    momentum_1h=0.8
)

print(f"Reflexivity: {score}")  # -7 to +7
print(f"Stage: {reflex.stage}")  # BUST → NEUTRAL → BOOM

if reflex.stage in ["LATE_BOOM", "BLOWOFF"]:
    # Momentum driving price — reduce exposure
    size *= 0.25

Based on Soros reflexivity theory. Detects self-reinforcing feedback loops.

8. SelfMetrics

Agent performance tracking.

from agentkit import SelfMetrics

metrics = SelfMetrics("metrics.json")

# Record every action
metrics.record_action(
    action="BUY",
    size=50.0,
    context={"signal": 5, "regime": "normal"}
)

# Record outcomes
metrics.record_outcome(pnl=12.50, fees=0.50)

# Get performance report
report = metrics.report()
print(f"Win rate: {report['win_rate']}%")
print(f"Sharpe: {report['sharpe_ratio']}")
print(f"Avg hold time: {report['avg_hold_hours']}h")
print(f"Best condition: {report['best_condition']}")

Design Philosophy

1. File-based persistence. JSON files, not databases. Your agent should run on a $6/month server, not require infrastructure.

2. Prompt-ready output. Every module has to_prompt_block(). The primitives are designed to inject context into LLM prompts.

3. Decay over deletion. Old lessons decay in weight but persist. Your agent's history is never erased, just de-prioritized.

4. Composable. Use one module or all eight. They don't depend on each other. Mix with any agent framework.

5. Observable. Every module has stats(). You can inspect your agent's learning at any time.

Installation

pip install bitbot-primitives

Or from source:

git clone https://github.com/johhnyg/agentkit
cd agentkit
pip install -e .

Requirements

• Python 3.9+
• No external dependencies (stdlib only)

Origin

These primitives were extracted from a live trading bot that has been running autonomously since February 2026. The bot manages real money, makes its own decisions, and learns from outcomes — with no human in the loop for weeks at a time.

Every pattern in agentkit exists because we hit a wall without it.

License

MIT

Contributing

Issues and PRs welcome at github.com/johhnyg/agentkit.