ABCI

Define the project once. Agents run it.

ABCI is an agent orchestration platform. You describe a software project as a dependency graph, and coding agents execute it end to end, with no engineer babysitting parallel sessions.

Project: add multi-currency checkout. 10 tasks across 9 microservices, each in its own repo.

0agent sessions in flight
0peak in flight
0 / 10tasks done
0sessions you tend with ABCI
Ten tasks waiting. Press Run project.

Without ABCI, every one of those sessions is yours to open, feed with context from the services upstream, and watch. With ABCI you define the graph and review the result.

Waiting on dependenciesAn agent is workingDone, downstream unblocked

The problem: one engineer, many agent sessions

Coding agents are good at a single task in a single repo. Real changes cross microservices: one feature touches pricing, cart, payments, the gateway and two clients, and each service needs its own session. The engineer becomes the scheduler and the message bus between them.

Today

  • Open one session per service repo and keep them all alive
  • Relay API contracts and schema changes from one session to the next
  • Track which service is blocked on which, and notice failures
  • Re-run downstream sessions whenever something upstream changes

With ABCI

  • Define the tasks, services and dependencies once
  • Tasks start automatically when their inputs are ready
  • Contracts and context pass downstream automatically
  • Failures are diagnosed and retried; you review the result

How it works

Define

Describe the project as tasks and dependencies. Claude can help break a high-level goal into well-scoped tasks.

Execute

A Claude-powered agent takes each task as soon as everything it depends on is finished. Independent tasks run in parallel.

Review

When the graph completes, you review the outputs instead of supervising each step along the way.

What Claude does in ABCI

Claude is the execution layer of the product.

Task execution
Writes, modifies and tests code for each node in the graph, using tool use inside the codebase.
Planning
Decomposes a high-level project description into a dependency graph of scoped tasks.
Handoffs
Summarizes the output of finished tasks so dependent tasks start with the context they need.
Recovery
Diagnoses failed tasks and retries them without waiting for a person.

Where we'd like support from Anthropic

  • API credits and priority rate limitsOur workload is many concurrent, long-running agent sessions per project, so throughput matters most.
  • Guidance on multi-agent workflowsContext handoffs, failure recovery, and cost optimization at scale.
  • Early accessAgentic and Claude Code features and model previews.
  • IntroductionsTeams building agent infrastructure who could become design partners.