lambdagent/docs/REAL_WORLD_DEFECTS.md

Real-World Agent Framework Defects: Evidence for lambdagentpaas Static Analysis

Date: 2026-04-04 Sources: GitHub Issues from Claude Code, LangChain, CrewAI, AutoGen, LangGraph Total Issues Analyzed: 40+ Conclusion: Every category of defect maps to a lambdagentpaas static guarantee

---

1. Cost Explosion ($342 Single Session — Claude Code)

anthropics/claude-code#38029 — Session Resume: $342.74 in One Session

• What happened: Session resume generated 652,069 output tokens without user input
• Cost: $342.74 in a single session
• Root cause: Resume logic auto-generated massive output with no budget check
• lambdagentpaas prevention:
  • Graded types (p, t, l, m): compile-time upper bound on tokens and cost
  • CEK Yield: cost check at every oracle call → budget breaker before exceeding $X

anthropics/claude-code#4095 — 1.67 BILLION Tokens in 5 Hours

• What happened: Catastrophic runaway event consumed 1.67 billion tokens
• Cost: Potentially thousands of dollars
• Root cause: No per-step cost tracking, no budget ceiling
• lambdagentpaas prevention:
  • Cost monotonicity (Paper II Prop. 23): c' ≥ c at every CEK step
  • Graded iteration g^n = (p^n, n·t, n·l, n·m): Loop cost upper bound computed at compile time

anthropics/claude-code#9704 — Orphaned Process: $50-100 Over 2 Days

• What happened: Shell process ran infinite loop for 2+ days, consuming ~2.88M tokens
• Cost: $50-100 completely undetected
• Root cause: No process lifecycle management, no effect isolation
• lambdagentpaas prevention:
  • Effect handlers: io effect requires handler → orphaned process impossible without handler leaking
  • Lint L004: no terminate condition detected at compile time

anthropics/claude-code#34629 / #41930 — Prompt Cache Bug: 10-20x Cost for 28 Days

• What happened: Cache hit rate dropped from 97% to 4%, silently inflating costs 10-20x
• Duration: 28 days across 20 versions, affecting all paid tiers
• Impact: Front-page Hacker News (136 points), The Register coverage
• lambdagentpaas prevention:
  • Graded types: Expected cost per call is part of the type → 10x deviation detectable as type violation
  • Effect annotation llm(m): Each model call carries expected cost grade → runtime deviation triggers alert

anthropics/claude-code#9579 — Autocompact Loop: 6x Token Spike

• What happened: Daily usage spiked from 17M to 96-108M tokens/day
• Root cause: Compaction triggers re-compaction in a cycle
• lambdagentpaas prevention:
  • Lint L004: Loop without effective terminate condition
  • Store independence: Compaction writes to same state it reads → violates writes(f) ∩ reads(f) ≠ ∅ cycle detection

---

2. Infinite Loops (7 Distinct Patterns — Claude Code Alone)

anthropics/claude-code#27281 — Agent Stuck: "Let me write" Without Writing

• What happened: Agent repeatedly stated "let me write the document" without ever calling the Write tool, consuming entire context window
• lambdagentpaas prevention:
  • Lint L004: ReAct loop where no branch leads to terminate
  • CEK trace: Continuation stack K shows no toolK frame ever pushed → detectable as non-progress

anthropics/claude-code#19699 — Same Failing Command Repeated Indefinitely

• What happened: Claude repeats identical failing command without modification, burning tokens
• lambdagentpaas prevention:
  • Store independence: state(step_n) = state(step_{n-1}) → identical state loop detected
  • Guard validator: P(output) = false for repeated identical errors → Guard retry limit enforced

anthropics/claude-code#6004 — Infinite Compaction Loop

• What happened: "Compacting conversation..." triggers again immediately after completing
• lambdagentpaas prevention:
  • Loop variant analysis: Compaction should reduce context size → if it doesn't, Y combinator base case unreachable
  • Lint L004a: TRUE semantic defect — loop body doesn't converge

anthropics/claude-code#13996 — No Kill Switch for Runaway Task Agents

• What happened: Task agent stuck in infinite loop, no way to terminate, must kill process
• lambdagentpaas prevention:
  • CEK Yield: Every oracle call is a Yield point → cancellation checked between steps
  • CEK pause/resume: ⟨C, E, K, σ, c⟩ serializable → graceful stop at any point

anthropics/claude-code#30016 — 20+ Minutes on Same SQL Injection Approach

• What happened: Playwright SQL injection loop, repeating same failing approach
• Cost: 20+ minutes of Opus-level token consumption
• lambdagentpaas prevention:
  • Store independence: Repeated identical tool inputs detected
  • Graded cost: 20 min × opus rate → cost prediction flags this as expensive before running

anthropics/claude-code#24585 — Opus 4.6 Explore/Thinking Loops

• What happened: Opus model gets stuck exploring files and "thinking" without productive action
• lambdagentpaas prevention:
  • CEK trace: K stack shows only compK(explore, ...) frames, never toolK(write/edit, ...)
  • Progress guard: No state change after N steps → Guard failure

anthropics/claude-code#10505 — Memory Leak: 28GB from Infinite Write Loop

• What happened: --continue caused infinite session state re-serialization, 27.87 GB memory
• lambdagentpaas prevention:
  • Effect annotation: state(session) effect with write-amplification detected
  • Lint L004: Recursive state write without convergence condition

---

3. Type/Format Mismatches (Crash at Runtime After Spending Money)

LangChain #17336 — LineList Expected Dict, Got Int

• What happened: MultiQueryRetriever output parser expects dict, receives int → TypeError
• Cost: All preceding LLM calls wasted
• lambdagentpaas prevention:
  • T-Compose: output(llm_chain) = Int, input(LineList) = Dict → Int ≮: Dict → compile error

LangChain #10997 — Chain Returns Dict, Tool Wrapper Expects String

• What happened: RetrievalQAWithSourcesChain returns dict, but agent tool wrapper expects string
• lambdagentpaas prevention:
  • T-Compose: output(chain) = Dict, input(tool_wrapper) = Str → Dict ≮: Str

LangChain #13101 — StructuredOutputParser Always Fails

• What happened: LLM returns triple-backticks format, parser regex expects single-backtick
• lambdagentpaas prevention:
  • T-Compose: output(LLM) = Str, input(parser) = Json(S) → schema validation at type level
  • Guard: Guard(llm >> parser, is_valid_json, k=3) with retry

AutoGen #391 — String Treated as Dict → AttributeError

• What happened: 'str' object has no attribute 'get' — string response treated as dictionary
• lambdagentpaas prevention:
  • T-Compose: typeof(response) = Str, but code calls .get() expecting Dict

LangGraph #6533 — Interrupt Resume Values Misrouted Between Tools

• What happened: Both tool interrupts get same ID → resume values delivered to wrong tool
• lambdagentpaas prevention:
  • Store independence: Two tools writing to same interrupt namespace → writes(tool_a) ∩ writes(tool_b) ≠ ∅

claude-agent-sdk-python#571 — StructuredOutput Wrapper Mismatch

• What happened: Agent wraps output in {"output": {...}}, schema validation expects bare object
• lambdagentpaas prevention:
  • Json(S) subtyping: Json(object({output: S})) ≮: Json(S) → caught by structural subtyping

---

4. Parallel Agent Data Corruption

LangChain Forum — Race Condition: Parallel Tool Responses Out of Order

• What happened: Multiple tool calls in parallel, responses arrive out of order, corrupting message sequence
• lambdagentpaas prevention:
  • Pair confluence (Paper II Prop. 30): Requires store independence writes(f) ∩ writes(g) = ∅
  • CEK PairLK/PairRK: Deterministic left-first-then-right evaluation prevents reordering

LangChain Forum — Multi-Agent Race Condition: Silent Data Loss

• What happened: Agent A and B both read state S, produce updates A' and B', one overwrites the other
• lambdagentpaas prevention:
  • Store independence check: writes(A) ∩ writes(B) ≠ ∅ → StoreConflictError at compile time
  • SharedMemory append-only: Paper I store typing Σ' ⊇ Σ — append-only prevents overwrite

---

5. Other Frameworks: Same Patterns

CrewAI #3847 — max_iter Does Not Actually Stop the Loop

• What happened: handle_max_iterations_exceeded prepares answer but it gets overwritten; loop continues
• lambdagentpaas prevention:
  • Theorem 5.4 (Bounded fixpoint termination): fix_n GUARANTEES termination in n steps
  • Lint L003: maxSteps=0 / ineffective iteration cap

CrewAI #737 — Tool Called But Result Never Returns to Agent

• What happened: Agent calls tool with same params forever; tool result never feeds back
• lambdagentpaas prevention:
  • CEK C-CompRet: Value MUST pop continuation and flow to next step — architectural guarantee
  • T-Compose: Tool output type must match agent input type for next iteration

AutoGen #108 — Blank Messages in Infinite Loop (GPT-4)

• What happened: Agents send blank messages back and forth, hit rate limits, burn money
• Root cause: is_termination_msg checks exact string "TERMINATE", LLM writes "(TERMINATE)"
• lambdagentpaas prevention:
  • Lint L004c: String-based termination is fragile — flagged as INFO
  • terminate = λx.x: In λA, terminate is a TOOL (structural), not a string match (heuristic)

AutoGen #1070 — No Way to Calculate Chat Session Cost

• What happened: Users cannot determine total cost of GroupChat session
• lambdagentpaas prevention:
  • Graded types: g_total = g_agent1 · g_agent2 · ... · g_agentN computed at compile time
  • CEK cost vector: c = (tokens, latency, cost) accumulated precisely per step

---

Summary: Mapping to lambdagentpaas Features

lambdagentpaas Feature	Real Issues Prevented	Example Issues
T-Compose type checking	9+ type mismatch crashes	LangChain #17336, #10997, #13101; AutoGen #391; claude-sdk #571
Graded cost prediction	10+ cost explosions	Claude Code #4095 ($1000s), #38029 ($342), #34629 (10-20x); AutoGen #1070
Lint L004 (no terminate)	7+ infinite loops	Claude Code #27281, #6004, #19699; AutoGen #108, #391; CrewAI #737
CEK Yield (pause/cancel)	3+ runaway agents	Claude Code #13996, #30016, #24585
Store independence	4+ data corruption	LangGraph #6533; LangChain Forum race conditions; Claude Code #9579
Effect handlers	2+ environment confusion	Claude Code #9704, #38029
CEK cost monotonicity	3+ undetected spending	Claude Code #4095, #9704, #9579

Total: 38+ real issues across 5 major projects, every single one mappable to a lambdagentpaas static guarantee.

---

The Killer Statistic

Claude Code #34629 alone: A prompt-cache regression silently inflated costs 10-20x for 28 days across 20 versions.

If Claude Code's agent pipeline had been compiled with graded types:

Expected: g_call = (0.99, 1500 tokens, 1.2s, $0.005)
Actual:   g_call = (0.99, 30000 tokens, 1.2s, $0.10)    ← 20x deviation

lambdagentpaas runtime check:
  ⚠️ COST ANOMALY at Yield(llm):
    Expected token cost: 1,500 (from graded type)
    Actual token cost:   30,000
    Deviation: 20x — exceeds 2x threshold
    Action: PAUSE execution, alert developer

This bug would have been caught on Day 1, Call 1 — not Day 28.