Agent difficulty states (FSM)

How ReasonBlocks classifies agent difficulty across FAST, NORMAL, SLOW, and SKIP states, and what each state means for model routing and E-trace injection.

ReasonBlocks tracks how hard your agent is working on each step using a finite state machine (FSM). After scoring a step, the FSM looks at the recent difficulty history and transitions between states that describe the agent’s current trajectory — coasting through easy steps, working normally, struggling with a hard problem, or appearing completely stuck. The current state gates model routing and controls whether E-trace injections are retrieved at all.

States

State	Meaning
INIT	Starting state before any step has been scored. Transitions to NORMAL on the first scored step.
FAST	The agent has produced consistently low-difficulty steps recently. E-traces are skipped entirely.
NORMAL	Default operating state. Scoring, monitors, and E1/E2/E3 injections all run as normal.
SLOW	The agent has produced consistently high-difficulty steps recently. A stronger model may be routed.
SKIP	The agent has been in high-difficulty territory for an extended window. Indicates a deeply stalled run.
END	Terminal state. Once reached, the FSM no longer transitions.

Transition rules

The FSM uses three thresholds and a hysteresis margin to decide when to transition. All values are configurable.

Parameter	Default	Purpose
fast_threshold	0.2	A step below this score is considered “easy”.
slow_threshold	0.6	A step above this score is considered “hard”.
skip_threshold	0.85	A step above this score is considered “very hard” for the SKIP gate.
hysteresis_margin	0.1	Buffer applied when leaving FAST or SLOW to prevent rapid oscillation.

From NORMAL:

• Transitions to FAST if the last 6 steps all scored below fast_threshold (0.2).
• Transitions to SLOW if the last 5 steps all scored above slow_threshold (0.6).
• Otherwise stays NORMAL.

From FAST:

• Transitions back to NORMAL if the current score exceeds fast_threshold + hysteresis_margin (0.3).
• Otherwise stays FAST.

From SLOW or SKIP:

• Transitions back to NORMAL if the current score drops below slow_threshold - hysteresis_margin (0.5).
• Transitions to SKIP if the last 35 steps all scored above skip_threshold (0.85).
• Otherwise stays SLOW.

The hysteresis margin prevents the FSM from rapidly flipping between states when scores hover near a threshold. A single easy step does not exit SLOW — the score must drop meaningfully below the boundary.

What each state does

FAST skips the entire E-trace pipeline. No pattern-store lookups, no embeddings, no injections beyond what is already cached. This keeps overhead minimal when the agent is making consistent progress. Monitors still run in FAST state because loop detection is most useful precisely when the agent is moving quickly.

NORMAL runs the full pipeline: monitors evaluate, E1 is queried if the monitor gate allows it, E2 retrieves pattern-level guidance, and E3 provides universal rules (on the first call).

SLOW runs the full pipeline and, if you have configured model_routing, routes the request to a stronger model. Monitor injection cooldowns are shorter in SLOW state (every 2 steps instead of 3) so guidance arrives more frequently.

SKIP behaves like SLOW but signals a deeply stalled run. It applies the same 2-step cooldown for monitor injections.

Configuring thresholds

Pass custom thresholds when constructing ReasonBlocks. Only the values you specify are overridden; the rest use their defaults.

from reasonblocks import ReasonBlocks

rb = ReasonBlocks(
    api_key="rb_live_...",
    fsm_thresholds={
        "fast_threshold": 0.15,
        "slow_threshold": 0.65,
        "skip_threshold": 0.90,
        "hysteresis_margin": 0.08,
    },
)

Configuring model routing

Map FSM state names to model identifiers. ReasonBlocks transparently swaps the model on the affected steps.

rb = ReasonBlocks(
    api_key="rb_live_...",
    model_routing={
        "SLOW": "anthropic/claude-opus-4-5",
        "FAST": "anthropic/claude-haiku-3-5",
    },
)

A common pattern is to route SLOW and SKIP to a stronger model and leave NORMAL and FAST on your default. This concentrates compute budget on the steps where it matters most.