Spaces:
Paused
Paused
File size: 33,305 Bytes
3f2f227 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 | """
Autonomous Agent v2 β CGAE Economic Actor
==========================================
Implements the v2 Autonomous Agent Architecture specification.
Separation of Cognition from Economy
-------------------------------------
The LLM handles task *execution*. Everything else β contract evaluation,
bidding strategy, robustness tracking, financial management β is deterministic
code. This makes the agent's economic behaviour inspectable without LLM
introspection, and keeps gas costs low.
Layers
------
PerceptionLayer β constraint / domain pass-rate learning
AccountingLayer β balance, exposure, reserves, burn-rate
PlanningLayer β EV / RAEV contract scoring + strategy delegation
ExecutionLayer β LLM call with constraint-aware prompts, self-verify, retry
Strategies (pluggable via StrategyInterface)
--------------------------------------------
GrowthStrategy β robustness-investment growth; the Theorem 2 agent
ConservativeStrategy β low-risk, low-utilisation; survives longest
OpportunisticStrategy β high-risk, max-reward; highest variance
SpecialistStrategy β domain-focused; improves pass rate in chosen domains
AdversarialStrategy β probes system limits; validates Proposition 2
Migration (Phase 1)
-------------------
Drop-in replacement for the bare LLMAgent + manual logic in live_runner.py.
The runner still handles contract posting, acceptance and Economy settlement.
AutonomousAgent.plan_task() β replaces random.choice(available_tasks)
AutonomousAgent.execute_task() β replaces llm_agent.execute_task() + retry
AutonomousAgent.update_state() β replaces inline robustness update logic
"""
from __future__ import annotations
import logging
import math
import random
import re
import time
from abc import ABC, abstractmethod
from dataclasses import dataclass, field
from typing import Any, Optional
from cgae_engine.gate import GateFunction, RobustnessVector, Tier, TierThresholds
logger = logging.getLogger(__name__)
# ---------------------------------------------------------------------------
# Data structures
# ---------------------------------------------------------------------------
@dataclass(frozen=True)
class AgentState:
"""Complete agent state snapshot passed to strategies each planning cycle."""
# Identity
agent_id: str
model_name: str
# Robustness
certified_robustness: RobustnessVector
effective_robustness: RobustnessVector # after temporal decay
certified_tier: Tier
effective_tier: Tier
binding_dimension: Optional[str] # "cc", "er", or "as"
gap_to_next_tier: dict # dim -> gap float
# Financial
balance: float
available_for_contracts: float
active_exposure: float
remaining_ceiling: float
burn_rate: float
rounds_until_insolvency: float
roi: float
# Performance history
constraint_pass_rates: dict # constraint_name -> float
domain_pass_rates: dict # domain -> float
total_contracts_completed: int
total_contracts_failed: int
win_rate: float
# Temporal
time_since_certification: float
spot_audit_probability: float
@dataclass(frozen=True)
class ScoredContract:
"""A contract that has been pre-evaluated by the Planning Layer."""
contract_id: str
task_id: str
min_tier: Tier
domain: str
constraint_types: list # list[str]
reward: float
penalty: float
deadline: float
difficulty: float
# Computed by PlanningLayer
estimated_pass_probability: float
estimated_token_cost: float
expected_value: float # p*R - (1-p)*P - cost
risk_premium: float # penaltyΒ² / (2 * balance)
risk_adjusted_ev: float # EV - risk_premium
@dataclass
class ExecutionResult:
"""Result of executing a task through the ExecutionLayer."""
output: str
token_usage: dict # input_tokens, output_tokens
token_cost_eth: float
latency_ms: float
retries_used: int
self_check_passed: bool
self_check_failures: list # constraint names that failed self-check
self_check_diagnostics: dict # name -> diagnostic string
@dataclass
class RobustnessInvestment:
"""An instruction to invest in a robustness dimension."""
dimension: str # "cc", "er", or "as"
budget: float # ETH to spend
# ---------------------------------------------------------------------------
# Strategy interface and concrete implementations
# ---------------------------------------------------------------------------
class StrategyInterface(ABC):
"""Pluggable decision policy for the Planning Layer."""
@abstractmethod
def rank_contracts(
self,
eligible: list, # list[ScoredContract]
state: AgentState,
) -> list: # ordered list[ScoredContract]
...
@abstractmethod
def should_invest_robustness(
self, state: AgentState
) -> Optional[RobustnessInvestment]:
...
@abstractmethod
def max_utilization(self) -> float:
"""Fraction of budget ceiling willing to commit."""
...
class GrowthStrategy(StrategyInterface):
"""
Invests in robustness to unlock higher tiers.
The Theorem 2 agent: robustness investment as profit strategy.
Accepts any positive RAEV contract, prefers higher tiers.
Invests when binding dimension is within ``invest_threshold`` of next tier
and projected ROI is positive within 20 rounds.
"""
RAEV_MIN = 0.0
INVEST_THRESHOLD = 0.07 # Invest when < 0.07 from next tier threshold
MAX_INVEST_FRACTION = 0.20 # Max 20% of balance per invest action
def rank_contracts(self, eligible, state):
# Sort by RAEV; break ties by preferring higher tiers
return sorted(
eligible,
key=lambda c: (c.risk_adjusted_ev + c.min_tier.value * 0.005),
reverse=True,
)
def should_invest_robustness(self, state: AgentState) -> Optional[RobustnessInvestment]:
if state.binding_dimension is None:
return None
gap = state.gap_to_next_tier.get(state.binding_dimension, 1.0)
if gap > self.INVEST_THRESHOLD:
return None
# Can we afford it?
budget = min(
state.available_for_contracts * self.MAX_INVEST_FRACTION,
state.balance * 0.10,
)
if budget < 0.005:
return None
# Rough ROI check: does the tier upgrade pay back within 20 rounds?
tier_idx = state.certified_tier.value
next_tier_reward_uplift = (tier_idx + 1) * 0.01 # Rough per-round uplift
if next_tier_reward_uplift * 20 > budget:
return RobustnessInvestment(dimension=state.binding_dimension, budget=budget)
return None
def max_utilization(self) -> float:
return 0.70
class ConservativeStrategy(StrategyInterface):
"""
Only high-certainty, low-penalty contracts.
Never invests in robustness. Survives the longest.
"""
RAEV_MIN = 0.002
MAX_DIFFICULTY = 0.5
def rank_contracts(self, eligible, state):
# Prefer lowest-penalty contracts above the RAEV minimum
passing = [c for c in eligible if c.risk_adjusted_ev >= self.RAEV_MIN
and c.difficulty <= self.MAX_DIFFICULTY]
return sorted(passing, key=lambda c: c.penalty)
def should_invest_robustness(self, state: AgentState) -> Optional[RobustnessInvestment]:
return None # Never invests
def max_utilization(self) -> float:
return 0.30
class OpportunisticStrategy(StrategyInterface):
"""
Max-reward, high-risk. Uses raw EV (not risk-adjusted).
Most likely to go insolvent; highest upside in good rounds.
"""
def rank_contracts(self, eligible, state):
return sorted(eligible, key=lambda c: c.expected_value, reverse=True)
def should_invest_robustness(self, state: AgentState) -> Optional[RobustnessInvestment]:
# Only if stuck at T0 β must reach T1 to earn anything
if state.certified_tier == Tier.T0 and state.available_for_contracts > 0.02:
dim = state.binding_dimension or "as"
return RobustnessInvestment(dimension=dim, budget=state.available_for_contracts * 0.30)
return None
def max_utilization(self) -> float:
return 0.90
class SpecialistStrategy(StrategyInterface):
"""
Domain-focused: only accepts contracts in its two best domains.
Higher RAEV threshold for unfamiliar territory.
Invests in constraint types where failure rate exceeds 30%.
"""
SPECIALIST_RAEV_MIN = 0.001
GENERALIST_RAEV_MIN = 0.010
NUM_SPECIALTY_DOMAINS = 2
FAILURE_INVEST_THRESHOLD = 0.30
def rank_contracts(self, eligible, state):
top_domains = sorted(
state.domain_pass_rates,
key=state.domain_pass_rates.get,
reverse=True,
)[:self.NUM_SPECIALTY_DOMAINS]
def score(c: ScoredContract) -> float:
in_specialty = c.domain in top_domains
thresh = self.SPECIALIST_RAEV_MIN if in_specialty else self.GENERALIST_RAEV_MIN
return c.risk_adjusted_ev if (in_specialty or c.risk_adjusted_ev > thresh) else -999
ranked = [c for c in eligible if score(c) > -998]
return sorted(ranked, key=score, reverse=True)
def should_invest_robustness(self, state: AgentState) -> Optional[RobustnessInvestment]:
# Find the constraint type with worst pass rate
if not state.constraint_pass_rates:
return None
worst_dim_name, worst_rate = min(
state.constraint_pass_rates.items(), key=lambda kv: kv[1]
)
if worst_rate > (1.0 - self.FAILURE_INVEST_THRESHOLD):
return None
# Map constraint family β robustness dimension
dim = "cc"
if any(k in worst_dim_name for k in ("keyword", "factual", "accuracy")):
dim = "er"
elif any(k in worst_dim_name for k in ("caveat", "harm", "refusal", "ethical")):
dim = "as"
budget = state.available_for_contracts * 0.15
if budget < 0.003:
return None
return RobustnessInvestment(dimension=dim, budget=budget)
def max_utilization(self) -> float:
return 0.50
class AdversarialStrategy(StrategyInterface):
"""
Probes system limits: selects borderline contracts and minimally invests
in AS (trying to game the ethical gate). Validates Proposition 2.
"""
def rank_contracts(self, eligible, state):
# Prefer contracts where self-check shows near-borderline probability
return sorted(
eligible,
key=lambda c: abs(c.estimated_pass_probability - 0.55),
)
def should_invest_robustness(self, state: AgentState) -> Optional[RobustnessInvestment]:
if state.binding_dimension == "as" and state.available_for_contracts > 0.01:
return RobustnessInvestment(dimension="as", budget=0.005)
return None
def max_utilization(self) -> float:
return 0.95
STRATEGY_MAP: dict[str, StrategyInterface] = {
"growth": GrowthStrategy(),
"conservative": ConservativeStrategy(),
"opportunistic": OpportunisticStrategy(),
"specialist": SpecialistStrategy(),
"adversarial": AdversarialStrategy(),
}
# ---------------------------------------------------------------------------
# Perception Layer
# ---------------------------------------------------------------------------
class PerceptionLayer:
"""
Tracks per-constraint and per-domain pass rates from task history.
Updated after every contract settlement via update_from_result().
"""
def __init__(self):
# Running history: name -> list[bool]
self._constraint_history: dict[str, list] = {}
self._domain_history: dict[str, list] = {}
@property
def constraint_pass_rates(self) -> dict:
return {
name: (sum(hist) / len(hist))
for name, hist in self._constraint_history.items()
if hist
}
@property
def domain_pass_rates(self) -> dict:
return {
domain: (sum(hist) / len(hist))
for domain, hist in self._domain_history.items()
if hist
}
def update_from_result(self, task: Any, verification: Any):
"""Call after each verification to update running pass rates."""
domain = getattr(task, "domain", "unknown")
self._domain_history.setdefault(domain, []).append(
bool(getattr(verification, "overall_pass", False))
)
for c in getattr(task, "constraints", []):
passed = c.name in getattr(verification, "constraints_passed", [])
self._domain_history.setdefault(f"constraint:{c.name}", [])
self._constraint_history.setdefault(c.name, []).append(passed)
def estimated_pass_prob(self, task: Any) -> float:
"""
Estimate pass probability for a task based on constraint and domain history.
Falls back to 0.65 when no history is available β modern LLMs pass
straightforward tasks at well above chance, so 0.5 systematically
underestimates EV and suppresses all task selection at startup.
"""
domain = getattr(task, "domain", "unknown")
domain_rate = self.domain_pass_rates.get(domain, 0.65)
constraints = getattr(task, "constraints", [])
if not constraints:
return domain_rate
rates = [self.constraint_pass_rates.get(c.name, 0.65) for c in constraints]
constraint_rate = math.prod(rates) if rates else 0.65
return (constraint_rate + domain_rate) / 2.0
# ---------------------------------------------------------------------------
# Accounting Layer
# ---------------------------------------------------------------------------
class AccountingLayer:
"""
Financial management with layered reserves.
Reserves (in priority order, all deducted before contract funds):
MINIMUM_RESERVE β hard floor; triggers SelfSuspend if breached
AUDIT_RESERVE β 1 full 4-dim audit cycle
(gas reserve is implicit in MINIMUM_RESERVE for off-chain simulation)
available_for_contracts = balance - active_exposure
- MINIMUM_RESERVE - AUDIT_RESERVE
"""
MINIMUM_RESERVE: float = 0.05 # ETH hard floor
AUDIT_RESERVE: float = 0.02 # ~4 dims Γ 0.005 ETH
MAX_UTILIZATION: float = 0.70 # Max fraction of ceiling to commit
def __init__(self, initial_balance: float):
self.balance: float = initial_balance
self.active_exposure: float = 0.0
self.cumulative_earned: float = 0.0
self.cumulative_spent: float = 0.0
self.cumulative_penalties: float = 0.0
self._burn_samples: list = [] # Recent ETH-per-round costs
@property
def available_for_contracts(self) -> float:
return max(
0.0,
self.balance
- self.active_exposure
- self.MINIMUM_RESERVE
- self.AUDIT_RESERVE,
)
@property
def roi(self) -> float:
spent = self.cumulative_spent + self.cumulative_penalties
if spent == 0:
return 0.0
return (self.cumulative_earned - spent) / spent
@property
def burn_rate(self) -> float:
if not self._burn_samples:
return 0.001 # Assume small storage cost until we have data
return sum(self._burn_samples[-10:]) / len(self._burn_samples[-10:])
@property
def rounds_until_insolvency(self) -> float:
br = self.burn_rate
if br <= 0:
return float("inf")
return max(0.0, (self.balance - self.MINIMUM_RESERVE) / br)
def can_afford(self, penalty: float, token_cost: float) -> bool:
"""Check whether accepting a contract keeps us solvent."""
new_exposure = self.active_exposure + penalty
headroom = self.balance - new_exposure - self.MINIMUM_RESERVE - self.AUDIT_RESERVE
return headroom >= token_cost
def record_round_cost(self, cost: float):
self._burn_samples.append(cost)
def sync_from_record(self, record: Any):
"""Sync from Economy AgentRecord (source of truth for balance)."""
self.balance = record.balance
self.cumulative_earned = record.total_earned
self.cumulative_spent = record.total_spent
self.cumulative_penalties = record.total_penalties
# ---------------------------------------------------------------------------
# Execution Layer
# ---------------------------------------------------------------------------
class ExecutionLayer:
"""
Executes tasks with:
1. Constraint-aware system prompt injection
2. Self-verification using the same checks the verifier will run
3. Retry loop (up to max_retries) when self-check detects failures
Self-check only covers algorithmic constraints (format, keywords, JSON).
Jury evaluation cannot be pre-checked β this is by design.
"""
def __init__(self, llm_agent: Any, self_verify: bool = True, max_retries: int = 2):
self.llm = llm_agent
self.self_verify = self_verify
self.max_retries = max_retries
def execute(self, task: Any, token_cost_fn) -> ExecutionResult:
"""
Execute a task end-to-end and return a structured result.
``token_cost_fn()`` is called with (model_name, in_tok, out_tok) to
compute ETH cost; the caller owns cost accounting.
"""
system_prompt = self._build_system_prompt(task)
user_prompt = task.prompt
tokens_in_before = self.llm.total_input_tokens
tokens_out_before = self.llm.total_output_tokens
start = time.time()
output = self.llm.execute_task(user_prompt, system_prompt)
retries = 0
self_check_result: dict = {"passed": True, "failures": [], "diagnostics": {}}
if self.self_verify:
self_check_result = self._self_check(task, output)
for attempt in range(self.max_retries):
if self_check_result["passed"]:
break
retries += 1
retry_prompt = self._build_retry_prompt(
user_prompt, self_check_result["failures"],
self_check_result["diagnostics"],
)
output = self.llm.execute_task(retry_prompt, system_prompt)
self_check_result = self._self_check(task, output)
latency_ms = (time.time() - start) * 1000
in_tok = self.llm.total_input_tokens - tokens_in_before
out_tok = self.llm.total_output_tokens - tokens_out_before
token_cost = token_cost_fn(self.llm.model_name, in_tok, out_tok)
return ExecutionResult(
output=output,
token_usage={"input": in_tok, "output": out_tok},
token_cost_eth=token_cost,
latency_ms=latency_ms,
retries_used=retries,
self_check_passed=self_check_result["passed"],
self_check_failures=self_check_result["failures"],
self_check_diagnostics=self_check_result["diagnostics"],
)
def _build_system_prompt(self, task: Any) -> str:
base = task.system_prompt or ""
if not task.constraints:
return base
lines = [
base,
"\n\n[CONSTRAINT REQUIREMENTS β you MUST satisfy ALL of the following]",
]
for c in task.constraints:
lines.append(f" β’ {c.name}: {c.description}")
return "\n".join(lines)
def _self_check(self, task: Any, output: str) -> dict:
"""Run algorithmic constraint checks identical to what the verifier will do."""
failures: list = []
diagnostics: dict = {}
for c in task.constraints:
try:
passed = c.check(output)
except Exception:
passed = True # Don't penalise unknown constraint types
if not passed:
failures.append(c.name)
diagnostics[c.name] = self._diagnose(c, output)
return {
"passed": len(failures) == 0,
"failures": failures,
"diagnostics": diagnostics,
}
@staticmethod
def _diagnose(constraint: Any, output: str) -> str:
name = constraint.name
if "word_count" in name:
count = len(output.split())
return f"Word count is {count}"
if "valid_json" in name:
return "Output is not valid JSON"
if "keyword" in name or "contain" in name:
desc = getattr(constraint, "description", "")
return f"Keyword check failed: {desc}"
if "section" in name:
return "Required section(s) missing from output"
return f"Constraint '{name}' not satisfied"
@staticmethod
def _build_retry_prompt(original: str, failures: list, diagnostics: dict) -> str:
diag_lines = "\n".join(
f" - {name}: {msg}" for name, msg in diagnostics.items()
)
return (
f"{original}\n\n"
f"[REVISION REQUIRED]\n"
f"Your previous response failed these constraints:\n"
f"{diag_lines}\n\n"
f"Please regenerate your response, fixing these issues while "
f"preserving the quality of your answer."
)
# ---------------------------------------------------------------------------
# Planning Layer
# ---------------------------------------------------------------------------
class PlanningLayer:
"""
Evaluates available tasks using EV / RAEV and delegates ranking to the
injected strategy. Also decides whether to invest in robustness.
"""
def __init__(self, strategy: StrategyInterface, token_cost_fn):
self.strategy = strategy
self._token_cost_fn = token_cost_fn # (model, in_tok, out_tok) -> float
def score_task(
self,
task: Any,
state: AgentState,
pass_prob: float,
) -> ScoredContract:
"""Score a single task and wrap it as a ScoredContract."""
# Token estimate scales with task tier: simpler tasks use fewer tokens.
# T1β200+100, T2β400+200, T3β600+300, T4+β800+400
tier_val = getattr(getattr(task, "tier", None), "value", 2)
in_tokens = max(200, min(800, 200 * tier_val))
out_tokens = max(100, min(400, 100 * tier_val))
est_token_cost = self._token_cost_fn(state.model_name, in_tokens, out_tokens)
reward = task.reward
penalty = task.penalty
ev = pass_prob * reward - (1.0 - pass_prob) * penalty - est_token_cost
# Risk premium: convex in penalty/balance β agents become risk-averse
# as penalties approach their balance (spec Eq)
balance = max(state.balance, 0.001) # avoid divide-by-zero
risk_prem = (penalty ** 2) / (2.0 * balance)
raev = ev - risk_prem
return ScoredContract(
contract_id="", # filled in by caller
task_id=task.task_id,
min_tier=task.tier,
domain=task.domain,
constraint_types=[c.name for c in task.constraints],
reward=reward,
penalty=penalty,
deadline=0.0,
difficulty=task.difficulty,
estimated_pass_probability=pass_prob,
estimated_token_cost=est_token_cost,
expected_value=ev,
risk_premium=risk_prem,
risk_adjusted_ev=raev,
)
def select_task(
self,
available_tasks: list,
state: AgentState,
perception: PerceptionLayer,
accounting: AccountingLayer,
) -> Optional[Any]:
"""
Return the best task to attempt, or None if nothing is worthwhile.
Safety checks run first (hard gates).
Then contract evaluation.
Then strategy ranking.
"""
# --- Safety checks --------------------------------------------------
if state.balance < AccountingLayer.MINIMUM_RESERVE:
logger.warning(
f"[{state.model_name}] balance {state.balance:.4f} below minimum "
f"reserve β suspending"
)
return None
# --- Score eligible tasks -------------------------------------------
ceiling = state.remaining_ceiling
utilisation_limit = ceiling * self.strategy.max_utilization()
scored: list = []
for task in available_tasks:
# Tier eligibility
if task.tier.value > state.effective_tier.value:
continue
# Budget eligibility (approximate β exact check in economy)
if task.penalty > utilisation_limit:
continue
if not accounting.can_afford(task.penalty, token_cost=0.01):
continue
pp = perception.estimated_pass_prob(task)
sc = self.score_task(task, state, pp)
scored.append((task, sc))
if not scored:
return None
# --- Strategy ranking -----------------------------------------------
ranked_scores = self.strategy.rank_contracts(
[sc for _, sc in scored], state
)
if not ranked_scores:
return None
# To avoid repetition, pick randomly from top N (e.g., top 3)
top_n = ranked_scores[:3]
selected_sc = random.choice(top_n)
top_id = selected_sc.task_id
for task, sc in scored:
if task.task_id == top_id:
if sc.risk_adjusted_ev > 0 or state.effective_tier == Tier.T0:
return task
return None
def investment_decision(self, state: AgentState) -> Optional[RobustnessInvestment]:
return self.strategy.should_invest_robustness(state)
# ---------------------------------------------------------------------------
# Autonomous Agent
# ---------------------------------------------------------------------------
class AutonomousAgent:
"""
v2 CGAE economic actor.
Wraps an LLMAgent and adds:
- Perception (constraint/domain pass-rate tracking)
- Accounting (reserves, burn-rate, insolvency prevention)
- Planning (EV/RAEV task selection, robustness investment decisions)
- Execution (constraint-aware prompts, self-verification, retry)
"""
def __init__(
self,
llm_agent: Any,
strategy: StrategyInterface,
token_cost_fn, # (model_name, in_tok, out_tok) -> float
self_verify: bool = True,
max_retries: int = 2,
):
self.llm = llm_agent
self.model_name: str = llm_agent.model_name
self.strategy = strategy
self.perception = PerceptionLayer()
self.accounting: Optional[AccountingLayer] = None # set in register()
self.execution = ExecutionLayer(llm_agent, self_verify=self_verify,
max_retries=max_retries)
self.planning = PlanningLayer(strategy, token_cost_fn)
self._token_cost_fn = token_cost_fn
# Set by economy on registration
self.agent_id: Optional[str] = None
# Metrics
self.self_check_catches: int = 0 # self-check prevented a failure
self.retry_successes: int = 0 # retry turned a failure into a pass
self.strategy_actions: dict = {}
def register(self, agent_id: str, initial_balance: float):
"""Call once after Economy.register_agent() to initialise accounting."""
self.agent_id = agent_id
self.accounting = AccountingLayer(initial_balance)
def build_state(self, record: Any, gate: GateFunction) -> AgentState:
"""
Construct an AgentState from an AgentRecord + gate details.
Called at the start of every planning cycle.
"""
self.accounting.sync_from_record(record)
r = record.current_robustness or RobustnessVector(0.3, 0.3, 0.25, 0.5)
gate_detail = gate.evaluate_with_detail(r)
tier = gate_detail["tier"]
ceiling = gate.budget_ceiling(tier)
total = record.contracts_completed + record.contracts_failed
win_rate = record.contracts_completed / max(1, total)
return AgentState(
agent_id=record.agent_id,
model_name=self.model_name,
certified_robustness=r,
effective_robustness=r, # decay applied externally by Economy
certified_tier=tier,
effective_tier=tier,
binding_dimension=gate_detail.get("binding_dimension"),
gap_to_next_tier={
"cc": gate_detail.get("gap_to_next_tier") or 0.0
if gate_detail.get("binding_dimension") == "cc" else 0.0,
"er": gate_detail.get("gap_to_next_tier") or 0.0
if gate_detail.get("binding_dimension") == "er" else 0.0,
"as": gate_detail.get("gap_to_next_tier") or 0.0
if gate_detail.get("binding_dimension") == "as" else 0.0,
},
balance=record.balance,
available_for_contracts=self.accounting.available_for_contracts,
active_exposure=self.accounting.active_exposure,
remaining_ceiling=max(0.0, ceiling - self.accounting.active_exposure),
burn_rate=self.accounting.burn_rate,
rounds_until_insolvency=self.accounting.rounds_until_insolvency,
roi=self.accounting.roi,
constraint_pass_rates=self.perception.constraint_pass_rates,
domain_pass_rates=self.perception.domain_pass_rates,
total_contracts_completed=record.contracts_completed,
total_contracts_failed=record.contracts_failed,
win_rate=win_rate,
time_since_certification=0.0, # computed externally if needed
spot_audit_probability=0.0,
)
def plan_task(
self,
available_tasks: list,
state: AgentState,
) -> Optional[Any]:
"""
Select the best task to attempt this round.
Returns None if nothing worthwhile or reserves too low.
"""
task = self.planning.select_task(
available_tasks, state, self.perception, self.accounting
)
action = "bid" if task else "idle"
self.strategy_actions[action] = self.strategy_actions.get(action, 0) + 1
return task
def execute_task(self, task: Any) -> ExecutionResult:
"""Execute a task with self-verification and retry."""
result = self.execution.execute(task, self._token_cost_fn)
# Track self-check performance
if not result.self_check_passed and result.retries_used > 0:
self.retry_successes += 1
if result.self_check_failures:
self.self_check_catches += 1
return result
def investment_decision(self, state: AgentState) -> Optional[RobustnessInvestment]:
"""Return a robustness investment if the strategy calls for it."""
inv = self.planning.investment_decision(state)
if inv:
self.strategy_actions["invest"] = self.strategy_actions.get("invest", 0) + 1
return inv
def update_state(self, task: Any, verification: Any, token_cost: float):
"""Update perception and accounting after a contract settles."""
self.perception.update_from_result(task, verification)
self.accounting.record_round_cost(token_cost)
def metrics_summary(self) -> dict:
return {
"model_name": self.model_name,
"strategy": type(self.strategy).__name__,
"self_check_catches": self.self_check_catches,
"retry_successes": self.retry_successes,
"self_check_catch_rate": (
self.self_check_catches
/ max(1, self.self_check_catches + self.retry_successes)
),
"strategy_actions": self.strategy_actions,
"constraint_pass_rates": self.perception.constraint_pass_rates,
"domain_pass_rates": self.perception.domain_pass_rates,
}
# ---------------------------------------------------------------------------
# Factory
# ---------------------------------------------------------------------------
def create_autonomous_agent(
llm_agent: Any,
strategy_name: str,
token_cost_fn,
self_verify: bool = True,
max_retries: int = 2,
) -> AutonomousAgent:
"""
Instantiate an AutonomousAgent with a named strategy.
strategy_name: "growth" | "conservative" | "opportunistic"
| "specialist" | "adversarial"
"""
strategy = STRATEGY_MAP.get(strategy_name)
if strategy is None:
raise ValueError(
f"Unknown strategy '{strategy_name}'. "
f"Choose from: {list(STRATEGY_MAP)}"
)
return AutonomousAgent(
llm_agent=llm_agent,
strategy=strategy,
token_cost_fn=token_cost_fn,
self_verify=self_verify,
max_retries=max_retries,
)
|