Every agent run,
made smarter.
ReasonBlocks catches failures mid-run, cuts tokens, and compounds intelligence across every agent you deploy.
Benchmarks — SWE-bench Pro
Toggling ReasonBlocks on the same agent loop delivers measurable lifts.
Same agent code, same prompts, same models. The only variable is whether ReasonBlocks is wrapping the client. Run on SWE-bench Pro.
Full methodology — token counts, latency measurement, agent loop — in the whitepaper.
Capabilities
Six capabilities. One runtime.
Each one targets a specific failure mode. None require changes to your agent code.
swipe to explore
Capabilities
Six capabilities.
One runtime.
Each one targets a specific failure mode. None require changes to your agent code.
- 01Reasoning reuse
Surfaces past solutions when a similar problem returns.
- 02Semantic file memory
Recalls what a file means without re-reading it.
- 03Loop detection
Catches doom loops mid-run and injects a redirect.
- 04Tool supervision
Spots redundant fetches before they compound on the bill.
- 05Context compression
Folds older turns into summaries so long horizons stay coherent.
- 06Reasoning-aware context pruningpreview
Reasoning- and logic-based pruning vs semantic token pruning.
Reasoning reuse
Semantic file memory
Loads AUTH_TOKEN via os.getenv, validates JWT, exposes authenticate() and refresh().
Parses .env and merges with defaults. Returns Config dataclass.
Mounts auth + rate-limit middleware on FastAPI app.
Loop detection
Tool supervision
Context compression
Reasoning-aware context pruning
Problem: A factory makes 12 widgets per hour. After the first 4 hours, output halves— but this throttle only applies on weekdays. Today is Saturday. The factory runs for 6 hours. How many widgets are produced?
Step 1: Base rate is 12 widgets per hour, before any adjustments.
Step 2: After 4 hours, the throttle would halve output to 6 per hour for the rest of the shift.
Step 3: Critical: the throttle only applies on weekdays. Today is Saturday, a weekend day.
Step 4: Since it's a weekend, the throttle does not activate. The factory runs at the full 12 per hour for all 6 hours.
Step 5: Total widgets = 12 × 6 = 72.
Answer: 72 widgets.
A short math problem with a catch. The small words only on weekdays and does not activate are what make the answer 72. Drop them and you get 60.
Integrations
Drops into the agent framework you already use.
from langchain.agents import create_agent
from langchain_anthropic import ChatAnthropic
from reasonblocks import ReasonBlocks
rb = ReasonBlocks(api_key="rb_live_...")
mw = rb.middleware(agent_name="bugfixer", task="Refactor the auth module")
agent = create_agent(
model=ChatAnthropic(model="claude-sonnet-4-5"),
tools=your_tools,
middleware=[mw],
)
with mw:
agent.invoke({"messages": [("user", "Refactor the auth module")]})
Other framework? Drop us a line.