AI & platform engineers From prototype to product
Turn agent experiments into systems you can ship.
Move from prompt notebooks to versioned specifications with explicit tools, policies, budgets, evaluations, and lineage.
Build agents that get better with evidence.
Promptsheon is an adaptive agent engineering platform for building, executing, evaluating, and continuously improving AI agents and multi-agent systems. Every run becomes evidence for the next better configuration.
Fastify + Next.js 16
Strands Agents SDK
Cedar policy engine
Runs on your laptop
promptsheon.local:3000/capabilities/refund-triage
canary
Manifest Evals Activity sha256:9c4f…a02b
# capability
name: refund-triage
version: 3
graph:
- classify: { model: claude-3-5 }
- retrieve: { tool: orders }
- decide: ~ reviewer policy v2 # +12 lines
- respond: { guardrail: pii-redact } # scrub list
# eval
suite: refund-suite
threshold: 0.92
scorers: [exact-match, llm-judge, regex]Score
0.94 ▲ 0.02
Latency p95
1.2s ▼ 0.3s
Drift
0.3% stable
Approved by m.chen · 4 min ago
0x9c…a02b View audit
01
Compose the system
Define agents, tools, policies, context, models, and budgets as one immutable specification.
02
Measure execution
Observe quality, cost, latency, context, tools, permissions, and failure modes across real work.
03
Promote what works
Compare candidates, validate mutations, and keep a reversible lineage for every change.
Who it is for
From promising agent to dependable product.
Promptsheon gives the people who build, run, and govern agent systems a shared path from first experiment to a safer, measurable release.
See the architecture Runtime & operations teams Reliable execution
Operate the graph with measurable control.
Route models, schedule dependencies, invoke tools, reuse safe results, and enforce resource limits while work is running.
Product, security & risk owners Evidence over intuition
Promote improvements with evidence, not optimism.
Compare outcomes, cost, latency, permissions, and drift, then keep approval history and a reversible path for every release.
1
release artifact
Prompts, tools, policies, and limits move together.
5
adaptive loop stages
Create, execute, observe, learn, promote.
0
telemetry by default
Your infrastructure, your data, your operating boundary.
1
clear rollback path
Return to a known content hash when needed.
The control plane
Build systems that learn how to improve.
Promptsheon combines architecture, execution, measurement, and resource governance into one adaptive runtime for agentic software.
Architect
Translate objectives into immutable agent and organization specifications: responsibilities, topology, capabilities, policies, and success criteria.
Operator
Schedule dependencies, route models, invoke tools, manage retries, reuse safe results, and materialize outcomes within resource budgets.
Evaluator + Governor
Measure quality and efficiency, discover what changed, enforce hard constraints, and promote only the mutations supported by evidence.
What you get
Six capabilities that ship together.
No bolt-ons. No beta flags. Every release of Promptsheon includes the whole stack — DAG editor, eval engine, audit chain, canary rollout, maker-checker, and the self-evolution loop.
Visual DAG editor
Compose agents, policies, tools, and memory into a graph. Live execution preview against real data, on your machine.
Releases with canary
Weighted traffic split. Live eval scores monitor canary. One-click rollback — hash-linked and signed.
Maker-checker approvals
The creator cannot approve their own release. The reason and voter are persisted with the audit row.
Content-addressed CAS
Every compiled manifest is hashed and stored by content. Same input, same hash, immovable identity.
Evaluation engine
Declarative datasets, pluggable scorers, regression gates, parallel runs. Compare models side by side.
Self-evolution loop
Watches live eval scores; on regression, re-plans the manifest and re-releases with cooldown.
Workflow
From execution to improvement.
Every execution becomes evidence. The platform uses that evidence to decide what to change, what to keep, and what the system can safely afford.
- 01
Create
Turn an objective into an immutable agent or organization specification with explicit policies and success criteria.
Output Content-addressed AgentSpec - 02
Execute
The Operator schedules the graph, routes models, invokes tools, and produces a trace within its budget.
Output Execution trace + artifacts - 03
Observe + evaluate
Measure outcome quality, tokens, context, latency, cost, retries, tools, permissions, and human intervention.
Output Evidence vector + decision - 04
Learn
Compare like workloads, trace causality through lineage, and store reusable patterns about what improves an agent.
Output Pattern + recommendation - 05
Mutate → validate → promote
Test a targeted change through replay, benchmarks, shadow traffic, or controlled rollout before promotion.
Output Candidate lineage + promotion
Evidence trail
The release carries its evidence.
Content
sha256:9c4f…a02b
The exact manifest that was evaluated and approved.
Decision
0.92 / 0.90 threshold
The release gate and the score that cleared it.
Recovery
v2 ready to restore
The previous active version stays available for rollback.
Trust & governance
Built for procurement, security, and risk.
Every capability has a reproducible history. Every release has an approver. Every audit row is hash-linked and publicly verifiable.
SOC 2 controls mapped
Access, change management, logging, monitoring.
SBOM published per release
CycloneDX 1.5 with hash, license, and provenance.
Incident response playbook
Documented rollback, evidence collection, disclosure.
Content-addressed
Same input, same artifact hash, same provenance.
Maker-checker
Approval workflow; creators cannot self-approve.
Hash-linked audit
Append-only chain. Public verification endpoint.
Rollback
Revert any release in one click. State is preserved.
Cedar policies
Declarative, testable authorisation for every API call.
Air-gap ready
Runs offline. No telemetry. No outbound network calls.
Comparison
Not a notebook. Not hosted SaaS.
Promptsheon sits between ad-hoc notebooks and hosted agent platforms. You own the infrastructure; we give you the engineering.
| Notebooks | Hosted SaaS | Promptsheon | |
|---|---|---|---|
| Visibility | — | partial | self-hosted, full |
| Audit chain | — | provider-side | hash-linked, verifiable |
| Maker-checker | — | optional | enforced |
| Canary rollout | — | beta | first-class |
| Rollback | manual | snapshot | one click |
| Cost model | time × seats | per-token | flat (Apache-2.0) |
| Vendor lock | low | high | none |
Pricing
One tier. Self-hosted. Apache-2.0.
No per-seat pricing. No cloud bill. Run it on your own metal, behind your own firewall. Pay only for the LLM tokens you actually use.
Recommended
Self-host
Free forever
Apache-2.0. Run anywhere you can run Node.
- All 76 REST endpoints
- DAG editor, eval engine, audit chain
- Multi-provider LLM
- No telemetry
- GitHub releases + signed tags
Enterprise
Talk to us tailored
Air-gapped installs, SSO, custom integrations.
- On-prem deployment playbook
- SSO via your IdP
- Custom audit retention
- Named engineering contact
Build agents that improve.
Set up Promptsheon on your own infrastructure. Configure your provider,
create your first AgentSpec, and turn execution evidence into a better candidate.
Ten minutes from
git clone to first canary.
$ git clone https://github.com/sachncs/promptsheon.git
$ cd promptsheon
$ pnpm install
$ cp .env.example .env
$ pnpm dev
▸ backend listening on http://localhost:8080
▸ frontend ready at http://localhost:3000
✓ capability refund-triage@v3 deployed · sha256:9c4f…a02b