Conflicts — Reasoning

Per-category conflict detail. Summary + index: CONFLICTS.md.

Critical 1 — R4 $\oplus$ R5

Type: Mutually Exclusive

ReAct interleaves reasoning and observation — it can adapt mid-task based on what it discovers. ReWOO plans all tool calls upfront and executes them without mid-run observation. These two are fundamentally incompatible for the same task:

• ReWOO assumes tool results are independent of each other. If tool call 2 should depend on the result of tool call 1, ReWOO produces wrong behavior because it has already planned both in advance.
• ReAct assumes you don't know what you'll need next until you see the current result. If tool calls are independent, ReAct wastes 5× more tokens doing what ReWOO does in two calls.

Resolution rule:

• Independent parallel lookups (search, retrieve, fetch from multiple sources) $\to$ R5 (ReWOO): 5× token efficiency
• Exploratory tasks where each step informs the next $\to$ R4 (ReAct): adaptability is worth the cost
• If in doubt at design time: prototype with R4 to understand the dependency structure; migrate to R5 once it's clear which calls are independent

Never do: Use R4 on a task where all sub-problems are provably independent. Use R5 on a task where sub-problems are sequential and dependent.

Critical 5 — R13 $\to$ V8

Type: Prerequisite Dependency

R13 (CodeAct) achieves its ~20pp accuracy advantage over JSON tool calls by executing arbitrary Python code. This is only safe inside a constrained execution environment. Without V8:

• LLM-generated code has full access to the host filesystem
• LLM-generated code can make arbitrary network requests
• A prompt injection (V6 concern) can generate and execute malicious code with the agent's full permissions
• A reasoning error can generate destructive code with no blast radius limit

Resolution rule: R13 without V8 is not a valid configuration in any production or shared environment. Treat this as a broken dependency, not a tradeoff.

Implementation: Docker containers (production), gVisor (high-security), or CodeSandbox/E2B (hosted sandbox) are the current implementation options.

Connection C — R17 $\sim$ prefix cache

Type: Composability Tension ($\sim$)

When R17 (Self-Consistency Voting) wraps R2 (Few-Shot CoT) with a static exemplar block, the exemplar block qualifies as a cacheable prefix (mechanism 5). But if N samples are dispatched sequentially over time exceeding the provider TTL (~5 minutes), later samples lose the cache hit and re-pay full prefill.

Resolution (O18 applies): Fan out all N samples simultaneously in parallel (O4 Parallelization). Do not dispatch them sequentially. The first sample pays the cache write; all subsequent parallel samples hit the cache. This converts the token cost of N samples from N × full_prefill to 1 × cache_write + (N-1) × cache_read.

Connection I — R7 $\sim$ R4

Type: Composability Tension ($\sim$)

Each R7 (Reflexion) retry is a full new R4 (ReAct) trajectory. The episodic memory buffer — containing N-1 prior critiques — is appended to each subsequent Actor call. Retry N's Actor call attends over a longer prefix than retry N-1 (mechanism 2: O(n²) attention cost). The retry cost is not N × per-task cost — it is strictly super-linear.

Example: For a base trajectory of 2,000 tokens and 3 critiques of 300 tokens each: Retry 1 pays O(2000²); Retry 2 pays O(2300²); Retry 3 pays O(2600²). Total: approximately 20–30% more than 3 × O(2000²).

Resolution: (1) Keep critiques compact — the Distiller pattern applied to critique outputs reduces the super-linear growth. (2) Cap retries aggressively — V9 Bounded Execution should account for the super-linear cost, not just count retries. (3) Clear the episodic buffer after convergence; do not carry it into the next independent task.

Reasoning vs Reasoning

Reasoning vs Orchestration