01 · VALIDATE
Find the surviving truth
Determine which completed work is still valid.
Selective recovery for AI agents and autonomous workflows
When one step breaks, why restart everything that already worked?
Consistency preserves what is still valid, repairs what failed, and continues the workflow without restarting everything.
Recover the broken part. Keep the work that is still valid.
Exact match
same final result as a clean full restart
Less replay
valid work was preserved instead of repeated
Live OpenHands
real model and external tool recovery
Pilots open
design-partner workflows wanted
Detailed measurements from the live recovery run are shown in Results below.
The killer demo
Watch the failure. Then watch what we refuse to throw away.
A tool-using agent has already completed useful work. SolarGrid and WindCore succeed. HydroLink fails late. The agent keeps going and reaches the wrong answer. A clean restart works but repeats valid work. Consistency reaches the same result without repeating valid work.
Measured scenario replay
OpenHands + Ollama · Qwen3 1.7B
Objective
idle
Plan
idle
SolarGrid
idle
WindCore
idle
HydroLink
idle
Analysis
idle
Response
idle
Contaminated answer
—
Recovered answer
—
Measured result
Same answer. Less repeated work.
Both strategies ended with exactly “HydroLink is the strongest project.” The difference was how much valid work they repeated to get there.
| Metric | Restart | Consistency | Reduction |
|---|---|---|---|
| Model calls | 3 | 2 | −33.3% |
| Prompt tokens | 1,961 | 223 | −88.6% |
| Total tokens | 2,043 | 236 | −88.4% |
| External tool calls | 3 | 1 | −66.7% |
| Recovery wall time | 11.261 s | 6.451 s | −42.7% |
The product
A recovery decision layer, not another workflow engine.
Your application keeps control of execution, tools, checkpoints, and domain logic. Consistency determines what is still valid, where recovery should begin, and what does not need to run again.
02 · RECOVER
Repair the smallest justified boundary
Preserve valid work and repair only what failed.
03 · CONTINUE
Regenerate dependents
Rerun only the work affected by the failure, then verify the final result.
Evidence beyond the pitch
Different failure modes. Same recovery principle.
Beyond the live demo, Consistency has also been tested across partial failures, process restarts, invalid persisted state, multi-step recovery, and distributed-state conflicts.
OpenHandsLive model + tool recovery with exact final-output equivalence.
OpenCMOFresh-process reconstruction and hidden persisted-artifact invalidity.
LangGraphSemantic invalidity after nominal graph completion.
Microsoft Agent FrameworkDAG recovery, multi-fault and state-authority experiments.
Current status
Technical alpha. Design-partner pilots open.
The core recovery runtime and developer-facing API are packaged. The next product milestone is external use on workflows people already care about.
Looking for one real workflow where restart is expensive
If your agent loses useful work after a failure, we want to see whether Consistency can preserve the valid work and reduce unnecessary recovery.
Pilot with Consistency
Your agent already knows how to do the work. It should not have to redo all of it.
Bring us a tool-using agent, long-running workflow, or autonomous process with partial-failure restart pain. We will measure selective recovery against the way you recover today.
A good pilot has one painful restart.
We are looking for workflows where useful state exists before a later failure forces broad replay, repeated model calls, repeated APIs, or manual recovery.
A tool or service can fail after earlier work succeeds.
Restart or replay repeats work you believe should still be valid.
You can define evidence for whether a stage is trustworthy.
Pilot request received. We’ll review the workflow and follow up using the email you submitted.
Oluseyi Christopher Olusola · Founder, Consistency