agentc
agentc build your_agent
Agents thatsurviveproduction.

agentc compiles declarative agent manifests into durable, production ready artifacts. Checkpointed graph execution means your agents recover from any failure, automatically, by design.

written in
Rust
status
closed alpha
license
MIT
agentc — build
cat my_agent/agent.acl
agent "assistant" {
archetype = "standalone"
model {
provider = "anthropic"
name = "claude-sonnet-4-6"
}
tool "web_search" {
source = "registry://agentc/web-search@v1"
}
}
agentc build ./my_agent
fetch web-search@v1
resolve standalone
codegen rust project
compile cargo build --release
./artifacts/build/my_agent 43.4 MB
./artifacts/build/my_agent
agent running on :8080

archetypes

One manifest.
Any deployment target.

input
agent "assistant" {
archetype = "standalone"
model { ... }
tool "search" { ... }
}
compiler
agentc
fetch·codegen·compile
output
./artifacts/build/my_agent 43.4 MB
standalone · linux/amd64
embedded runtime · http ready
standalone
Standalone Binary

A single cross-platform executable. Copy it to any server and run. Zero runtime dependencies.

-> ./artifacts/build/my_agent
serverlesssoon
Serverless Function

Compiled for Lambda, Cloudflare Workers, or other FaaS targets. Cold-start optimized, event-driven.

-> ./artifacts/build/my_agent.zip
agentc cloudsoon
Managed Deploy

Compile directly for agentc Cloud infrastructure. One command from manifest to live endpoint.

-> agentc.sh/deploy

patterns

Two ways to put
agents to work.

01

Agent Interface

Your agent is the product. It handles the full conversation loop, calls tools on behalf of the user, and streams responses over HTTP to the UI. Compile once, serve anywhere.

customer supportcopilotsassistantssearch
POST /v1/run HTTP 200live

Summarize last quarter's sales data and flag any anomalies.

a

calling tool

query_database("sales_q4_2025")

Q4 revenue was $2.4M, up 18% YoY. Anomaly detected: Nov 14 spike (+340% single day).

my_agent run "summarize sales data"completed in 7.1s
09:14:02inforun startedrun_id=a3f9c2
09:14:02infotext message contentLet's summarize the sales data.
09:14:03toolfetch_datasales_q4_2025
09:14:05toolrun_analysisQ4 2025 + 18%
09:14:06tooldetect_anomaliesNov 14 spike (+340%)
09:14:09toolgenerate_reportsummary.txt
09:14:09inforun finisheddata=[...] anomalies=[Nov 14]
task complete
02

Agent Operator

Your agent runs entirely in the background. It receives tasks from a queue or system event, executes the task with reasoning and tools, and emits structured output. No user interaction required.

automationpipelinesworkflowsdata processing

runtime

Your agent runs.
Whatever happens.

Every compiled agent is built on a graph runtime. Nodes execute in sequence, each one checkpointed. If anything goes wrong, the next run resumes from the last saved state.

durable
Crash recovery

Every node writes a checkpoint. A crash, restart, or context window reset resumes from the last safe point with no data loss and no manual retry logic.

observable
Full trace visibility

Model calls, tool invocations, and graph transitions are all traced. Emit to any OpenTelemetry-compatible backend with zero additional configuration.

scalable
Horizontal scale

Agents are stateless between checkpoints. Run as many instances as needed behind a load balancer with no shared state and no coordination overhead.

ecosystem

Tools,
your way.

custom tools

Write tools in TypeScript or Python. The compiler bundles and links them into the output artifact.

TypeScriptPythonRustsoonGosoon
mcp tools

Any MCP-compatible server works out of the box. Point your manifest at it and the compiler handles the rest.

FilesystemDatabasesBrowserAny MCP server
tool registry

Reference community and private tools by URI. The compiler fetches and links them at build time.

registry://agentc/web-search@v1registry://agentc/code-exec@v2oci://my-org/private-tool@sha256:...

orchestration

Agents that work
together, automatically.

Agents deployed on the same network discover each other via mDNS and expose themselves as callable tools. An orchestrator can delegate subtasks to specialists in parallel, then synthesize the results.

discovery
Zero config

Agents on the same LAN announce themselves automatically. No registry, no service mesh, no manual wiring.

delegation
Agents as tools

Any deployed agent is callable as a tool from any other agent in the network.

parallel
Parallel execution

Subtasks are dispatched concurrently across specialist agents for faster end-to-end completion.

context
Context isolation

Each agent manages its own context window, keeping individual tasks focused and efficient.

early access

Ship your first
agent today.

Get the compiler, the runtime, and a direct path from manifest to production binary.