The Graff harness is out. 83 ± 3 on Terminal Bench. State of the art.
github →The self-improving coding harness for AI agents
83%
Terminal Bench (±3)
harness eval · May 2026
48ms
cold start p50
create + first exec, intra-datacenter
21x
faster than fork/exec
0.7ms in-process vs 15ms spawn
100%
task isolation
per-run ephemeral environments
Sandbox cold start / ms to first command
intra-datacenter / Apr 2026 / 10 trials each
Pay as you go
Instant sandboxes. Pay as you go.
Spin up sandboxes in milliseconds, shut them down just as fast. Usage-based pricing, billed per second. Volume discounts as you scale.
| Resource | Unit | Per second | Per hour |
|---|---|---|---|
| Compute | vCPU | $0.0000196 | $0.071 |
| Memory | GiB | $0.0000063 | $0.023 |
| Storage | GiB | $0.00000004 | $0.0002 |
* Storage: 5 GB free, then billed per GiB/s. All prices in USD.
Why it's different
Cold start matters more than throughput
Agents iterate in tight loops. A 1s sandbox startup compounds into minutes of dead time per task. At 48ms, agents feel instant.
Isolation without overhead
Every agent run gets a clean environment. No state bleed between tasks, no permission sprawl. Tear-down is automatic.
Tuned for terminal tasks
Terminal Bench measures what matters: file I/O, shell commands, grep, patch, compile. Not synthetic throughput. Real agent work.
Terminal Bench
What it measures. Real shell tasks: write a file, run a test, apply a patch, search a codebase. The benchmark an agent actually has to pass.
Our score.83 ± 3 on Terminal Bench. Ahead of every published harness we've seen. Full run is on GitHub.
Why it matters. Throughput benchmarks hide latency. Terminal Bench surfaces the pause an agent actually feels between steps.
Web crawl sandbox
The same sandbox can crawl almost any public webpage: docs, marketing sites, dashboards, forms, changelogs, and JS-heavy apps. Auth walls, robots rules, and rate limits are respected.
agent sandbox crawl https://docs.example.com
render page - extract links - queue useful pages - return map
Open a clean browser context
A fresh sandbox navigates to the target URL without sharing cookies, files, or state.
Render and read the page
Kuri captures text, links, forms, screenshots, and stable action refs from the loaded page.
Follow the useful links
The agent queues same-site pages, docs, pricing, changelogs, forms, and API calls instead of crawling noise.
Return an inspectable map
You get markdown, screenshots, network traces, and the exact actions the agent took.
Agent-to-agent chats
The next thing landing on top of the sandbox: a structured chat plane between agents. One planner, many workers, every exchange typed and replayable. The whole thing sits on the same 48ms sandbox runtime you already get today.
planner -> worker[crawl]: { intent: "summarise docs", url: "..." }
worker[crawl] -> planner: { status: "done", refs: [...], summary: "..." }
planner -> worker[edit]: { intent: "apply patch", file: "..." }
Agents that talk to each other
A planner agent breaks a task into subtasks and dispatches them to specialist agents. Each one runs in its own sandbox, reports back over a typed message channel, and the planner re-plans on partial results.
Typed message channel
Not raw stdout. Every message is a structured envelope: sender, role, intent, payload, citations. The planner can route, retry, or fan out without parsing freeform text.
Inspectable transcripts
Every conversation is a durable, replayable trail. Step through the exchange, fork from any point, or rerun with a different model. No more guessing why an agent gave up.
Per-agent sandboxes
Each participant lives in its own 48ms sandbox: clean filesystem, scoped writes, isolated network. A worker agent crashing or going off-script never leaks state into the planner.
Same waitlist. Early access flips on for sandboxes first, then chats. Drop your email below.
Specialist agents
Same 48ms sandbox, role-tuned for the work. Each ships with a curated tool surface, prompts, and evals. Drop one in over MCP.
Design agent
SoonShip UI faster than your designer can comp it.
From a one-line prompt to a production-ready component. Reads your design tokens, picks accessible color, and iterates against real preview screenshots until the diff matches the spec.
- Tailwind / shadcn aware
- Screenshot-driven iteration
- Figma import (preview)
Finance agent
SoonAn analyst that doesn't sleep through close.
Reads ledgers, reconciles books, runs scenario models, and writes the variance commentary. Ships PDF and Excel, and cites every number back to its source row.
- Spreadsheet + DB native
- Cited variance reports
- SOX-friendly audit trail
Early access
Join the waitlist
We're rolling out in batches. Drop your email and we'll reach out when your slot opens.
By joining you agree to receive product updates and marketing emails from Codegraff. We'll never share your address. You can unsubscribe at any time. See our Privacy Policy.