Plan as a graph
Tasks, their dependencies and the tool each needs. The bench finds the critical path and how many agents actually help.
An open workbench for AI agents
Give every agent a bench, a toolbox and a memory you can audit.
Sporebench is where you plan what an agent will do, decide which tools it may touch, price the run before you start it, and seal its memory with hashes anyone can check. Permission policies can be anchored on Solana mainnet from your own wallet.
CA
0 tools granted in experiment 02
Bench conditions Solana mainnet, read live
- Slot
- —
- Throughput
- — tx/s
- Priority fee, median
- — µ◎/CU
- Epoch
- —
- SOL
- —
Agents are only as trustworthy as their bench.
Permissions you can fingerprint
A matrix of agents and tools becomes a canonical policy with a SHA-256 fingerprint you can sign or anchor on mainnet.
Memory with receipts
Notes are hashed into a Merkle tree, so you can prove a single memory was there without showing the rest.
Experiment 01 · Task graph
Lay out the work before an agent touches it.
- Makespan
- —
- Critical path
- —
- Total work
- —
- Speed-up from parallel agents
- —
Experiment 02 · Tool permissions
Decide what each agent may touch, then seal it.
Click a cell to cycle off, ask first and auto. The bench scores each agent's blast radius and checks for the dangerous three-way combination.
Policy fingerprint SHA-256
computing…
Signing is free and stays off chain. Anchoring writes one SPL Memo from your wallet.
Experiment 03 · Run cost
Price a run before you press go.
| Model | In $/M | Out $/M | Per run | Per month | Relative cost |
|---|
Cheapest option per month in SOL at —: —
Experiment 04 · Memory seal
Seal an agent's memory, prove one note at a time.
0 notes · about 0 tokens estimate, 4 chars ≈ 1 token
Root—
Experiment 05 · Agent account check
Is this agent's Solana account fit for work?
Agents that pay for receipts need an account with fees on hand and no clutter. Paste the address an agent will use, or connect your wallet.
Results appear here: balance, token accounts, reclaimable rent and recent activity.
Observation · one run, start to finish
What a well-kept run looks like.
- Request A message arrives in the inbox: "audit wallet and write a note".
- Plan The agent lays the work out as a graph and checks it against its policy.
- Tools Only granted tools fire. Anything marked "ask first" pauses for you.
- Memory Results become notes, each one hashed into the memory tree.
- Receipt The new root and the policy fingerprint are written into one memo.
How the bench works
- i.
Sketch the work
Break a job into tasks with durations, the tool each one needs, and what it waits for. The bench sorts it, finds the critical path and schedules it across as many agents as you allow.
- ii.
Grant tools on purpose
Every agent gets a row of permissions. The policy is serialised in a fixed order and hashed, so the same policy always has the same fingerprint.
- iii.
Price, then run
Steps × tokens × list price, with prompt caching, gives a cost per run and per month. The SOL figure uses the live Jupiter price.
- iv.
Seal and anchor
Memory notes become a Merkle root. The root and the policy fingerprint can be signed by your wallet or written to mainnet as an SPL Memo you pay for and control.
Ledger of sources
| On this page | Kind | Source |
|---|---|---|
| Bench conditions | live | Solana mainnet RPC: getSlot, getRecentPerformanceSamples, getRecentPrioritizationFees, getEpochInfo |
| SOL price | live | Jupiter price API v3 |
| Task graph and schedule | computed | Your inputs, sorted and scheduled in the browser |
| Permission risk score | model | Our weights per tool and combination, shown in the FAQ |
| Policy fingerprint, memory root, proofs | computed | SHA-256 via Web Crypto |
| Model prices | entered | List prices typed in by us; editable, not fetched |
| Agent account check | live | Solana mainnet RPC through a read-only proxy |
| Anchor receipt | on chain | SPL Memo signed and paid by your own wallet |
FAQ
Does Sporebench run my agents?
Not on this page. This is the planning and audit bench: it computes plans, policies, costs and memory proofs in your browser. Nothing you type is sent to a server, except addresses you check, which go to Solana mainnet through our read-only RPC proxy.
What exactly gets written on chain when I anchor?
One SPL Memo instruction, signed by your wallet, with the text sporebench:policy:v1: followed by the SHA-256 fingerprint and, if present, the memory root. It sends no SOL or tokens anywhere. You pay the network fee, which we estimate with getFeeForMessage before your wallet opens. It is permanent and public.
How is the risk score calculated?
Each tool has a weight (files.write 3, shell 5, web.fetch 2, wallet.sign 6, email.send 3, and so on). "Ask first" counts at 40% and "auto" at 100%. An agent that has private data, untrusted input and a way to send data out gets a +8 penalty and a warning. It is a model to make you think, not a security audit.
Why does the fingerprint stay the same when I refresh?
The policy is serialised with agents and tools in a fixed order, so equal policies produce equal bytes and therefore equal SHA-256 hashes. Change one cell and the fingerprint changes completely.
Can I trust the model prices?
They are list prices we entered by hand and may be out of date. Every price cell is editable. The arithmetic is exact; the inputs are yours to correct.
Does the site ever hold funds or keys?
No. Connecting a wallet shares only your public key. Signing happens inside your wallet after you approve it, and the only transaction this site can build is a memo from you to no one.
Is there a token?
If a Sporebench coin exists, its contract address will be shown at the top of the page and in the footer. Nothing on this page depends on it, and no feature is gated by it.