Abada Platform documentation

curl -fsSL https://install.abadaplatform.com/install.sh | bash

The transactional ACID rail

PostgreSQL is authoritative. Lock, validate, advance, persist, write history, facts and outbox records, then commit as one workflow-state transaction.

Studio is the authoring surface

Canvas and APL YAML are one document, the engine validates while you edit, Dry Run stays local and mocked, and Deploy & Start is the durable action.

Agents are durable work

Agent nodes compile to external tasks on the worker protocol. The engine never invokes a model inside a workflow transaction.

Agents act under policy

Agents call the tools of a governed MCP tool server: reads run freely, writes are journaled before they run, and an approval-required call waits for a person. A crash never repeats a write blindly. New in 1.1.0-rc.2.

Processes compose

A step can call another process, and an agent can delegate a sub-case to one, with only declared outputs coming back and lineage linking every child. New in 1.1.0-rc.2.

Optimization is governed

The Insight Engine turns terminal facts into validated APL proposals that reach production only through explicit human review. No auto-apply.

Agents advise, rules decide, humans approve, PostgreSQL remembers

Abada runs business processes in which AI does real work under rules you can audit:

• APL is the process language. Studio draws abada.io/v1 YAML as a diagram, and the engine compiles it directly into the executable graph. BPMN 2.0 files import and run on the same engine, extended with Abada’s agent and review steps.
• The engine enforces the contract. Every agent answer is checked against its schema and confidence threshold before it enters the process; rules and decision tables run deterministically inside the transaction.
• Real process shapes. Steps can repeat with a limit, route around errors, timeouts and weak AI answers, wait for a human decision with a required comment, escalate when late, and fall back to another model when a provider is rate-limited.
• Agents that act. Agents use tools from a project’s MCP tool servers, under read, write or approval-required policies; a person approves each risky call with its exact arguments; turns, tokens and spend are bounded; a routing agent chooses among declared routes, and an agent can delegate a sub-case to another governed process.
• Evidence for every decision. Each agent attempt, decision, route and tool call is in the audit trail — without prompts, secrets or comment text — and its cost is computed by the engine.

Start here

ConceptsEvery capability in a few lines: what it is, what the engine guarantees, where it stops.

QuickstartInstall the platform on one machine.

Tutorial: a reviewed AI replyBuild a complete governed AI process in Studio, step by step, with screenshots.

Tutorial: an agent that actsA refund agent with tools, an approved write, a crash that is not repeated and its cost.

Use casesFive runnable processes and what each one shows.

Loops, routes and review decisionsEvery shape a real process needs.

ArchitectureHow the transactional core keeps AI work safe.

Documentation tracks

1. User guide — install, build and run processes, operate them, and deploy the platform for a team.
2. Architecture — the transactional core, APL, agent execution, Insight, deployment, clustering and security.
3. Compatibility & migration — BPMN import, dialects and the supported subset.
4. Developer guide — the repository, engine and contract changes, and the evidence each change needs.

Release status

These guides describe the current dev tree.

• 1.1.0-rc.2 is the current evaluation release candidate (milestone M3, agents that act): tool servers and tool policies, tool approvals, the step journal and crash-safe writes, agent limits, evidence and cost, routing agents, call-process and agent delegation, live instance updates and the step inspector. See the release notes.
• 1.1.0-rc.1 (milestone M2, real process shapes): bounded loops, error and timeout routes, enforced service levels, model fallback, review outcomes with a required comment and incidents you can retry from Studio.
• 1.0.0-rc.8 remains available. PostgreSQL is the source of truth, API v1 and worker protocol v1 are frozen, and native APL and the BPMN subset are certified runtime surfaces.

The authoritative gates live in the roadmap and contracts.