Home / Blog / Bitter lessons we learned from building an open-source…
Tech News

Bitter lessons we learned from building an open-source coding agent

Imbue open-sourced Sculptor in mid-2026 as a customizable, model-agnostic coding environment that runs parallel agents inside local Docker containers.

By Dillip Chowdary • Nov 18, 2026 • Source: ideas.imbue.com

Bitter lessons we learned from building an open-source coding agent
Image: ideas.imbue.com

At a glance

Product name
Sculptor
Developer
Imbue
Initial release
2025
Open source release
Mid-2026
Core isolation
Docker containers
Primary supported models
Claude and Pi

AI research laboratory Imbue launched Sculptor, an open-source parallel-agent coding environment designed to coordinate autonomous programming models within customizable local developer workspaces. The release provides a containerized alternative to closed developer platforms like Cursor and Claude Code by pairing an extensible interface with multi-agent orchestration, as detailed in ideas.imbue.com's report on the project's development history. By executing autonomous agents locally rather than on remote hosts, the software provides developers with immediate control over repository changes across concurrent programming tasks.

This report examines the iterative engineering decisions, abandoned prototypes, and technical architecture behind Sculptor, alongside the practical operational patterns required to manage parallel autonomous coding agents. Engineering teams evaluating agentic development environments, local workflow automation, and open-source tooling can evaluate how container isolation and model-agnostic architectures alter day-to-day software construction.

01Hardware jailbreaks to software agents

The project originated from personal hardware tinkering, starting with childhood iPod jailbreaking experiments to run gaming emulators before transitioning into high school software development. The advent of modern large language models exposed an opportunity to radically accelerate software authoring, allowing individual developers to quickly construct applications that previously required substantial engineering investments.

Early agentic development began when GPT-4 API access first arrived, prompting experimentation with the ReAct framework to equip models with bash execution and terminal tooling. Those preliminary attempts struggled because early models lacked the reasoning capacity needed to reliably navigate terminal output and execute complex file manipulations.

02Tool scaffolding and the bitter lesson

Imbue CTO Josh and the engineering team initially responded to early model weaknesses by building Crafty, a prototype that generated parallel code diffs directly from API queries. Crafty focused on developer concurrency, allowing engineers to dispatch generation tasks to the model while continuing work on separate application components rather than watching single generations stream.

The tool soon succumbed to a classic machine learning reality where external engineering scaffolding is rendered obsolete by general model advances. Hand-crafted tooling built to help the model resolve linter errors became useless once updated base models resolved those syntax discrepancies independently without specialized prompt harnesses.

| Prototype | Design premise | Architecture | Observed failure mode | | Crafty | Parallel diff generation | Direct GPT API calls | Tooling replaced by native model capabilities | | Blackberry | Pre-written commit sequences | Reactive commit chains | Unnatural workflow lacking conversational chat | | Sculptor | Extensible parallel agents | Isolated Docker containers | Requires individual interface customization |

03Reactive commits and swarm week experiments

The team subsequently designed Blackberry, an experimental system named in reference to OpenAI's internal Strawberry project, that restructured development around predictive commit planning. Instead of writing application code directly, developers specified feature roadmaps as ordered hierarchies of descriptive git commits that the system executed sequentially.

Advertisement

Tech Pulse Daily

Get tomorrow's pulse first

Join engineers who read Tech Pulse before stand-up. Free, weekday mornings.

Blackberry featured a reactive propagation layer where edits applied to earlier commits automatically prompted downstream tasks to update their implementation files. Imbue developed the prototype during an intensive team swarm week involving engineers Josh, Johnny, Amy, Guinness, and Sam working collaboratively in a single room.

The team ultimately discarded Blackberry because writing prospective commit sequences proved unnatural for human engineers accustomed to design documents and conversational chat interfaces. Furthermore, advancements in model reasoning enabled frontier models to generate structured execution plans autonomously, eliminating the need for rigid human-engineered task breakdowns.

04Isolated containers and local developer pairing

Sculptor emerged as an extensible agentic development environment inspired by the minimalist design philosophy of the pi agent architecture. Rather than enforcing an inflexible remote operational paradigm like early iterations of Codex, Sculptor operates as a responsive desktop application running on local workstations.

  1. Workspace configuration Select the target repository and the base branch to fork within the environment.
  2. Agent assignment Designate the specific language model agent to execute tasks inside an isolated container.
  3. Parallel execution Dispatch background programming tasks across separate containers without manual worktree supervision.
  4. Pairing synchronization Merge selected container changes directly back into the local integrated development environment.

The system prevents accidental system corruption by encapsulating each active Claude agent within dedicated Docker containers running on the host system. A native pairing mode allows developers to inspect container workspaces and synchronize agent branch states directly into their local editor without managing git worktree collisions manually.

Note: Imbue launched Sculptor internally in 2025 before publishing the code under an open-source license in mid-2026, while ongoing engineering focuses on persistent panel layout configurations.

05Open source release and model agnosticism

The default Sculptor interface incorporates split workspace panels containing source code editors, interactive terminals, an active agent registry, chat panels, and integrated extensions for web browsing and task tracking. Imbue engineers routinely deploy Sculptor against its own codebase to identify workflow friction, repair internal defects, and accelerate platform development through recursive modification loops.

To accommodate divergent developer preferences, the platform maintains a model-agnostic runtime that allows engineers to execute models such as Claude and Pi side by side within identical project trees. Team members frequently reconfigure their interface layouts, prompting ongoing work on workspace profiles that persist customized panel configurations across separate development sessions.

Imbue established Sculptor as open-source software to ensure advanced developer infrastructure remains accessible as public utility tooling across the wider programming ecosystem. By lowering the operational overhead associated with multi-agent orchestration, the environment enables individual programmers to direct concurrent coding assistants through tailored local interfaces.

Frequently asked questions

What is Sculptor?

Sculptor is an open-source development environment built by Imbue that coordinates multiple parallel coding agents inside isolated Docker containers. It allows developers to customize panel layouts and run different language models side by side.

How does Sculptor differ from remote coding systems like Codex?

While early versions of Codex emphasized remote infrastructure with setup friction for long-running jobs, Sculptor functions as a tactile desktop application running on local machines. Developers run agents in local Docker containers and sync code via pairing mode.

Why did Imbue abandon the Blackberry prototype?

Imbue discarded Blackberry because engineers found pre-planning commit sequences unnatural compared to writing specs or using chat interfaces. Rapid model improvements in autonomous planning also rendered hand-specified commit scaffolding redundant.

Which models does Sculptor support?

Sculptor is built with a model-agnostic architecture inspired by the minimal core of the pi agent. It enables developers to run models like Claude and Pi simultaneously across different containerized tasks.

Sources

Dillip Chowdary

Author

Dillip Chowdary

Writes Tech Bytes coverage of AI, engineering, and the tools that actually ship. Editor of Tech Pulse Daily.

Related on Tech Bytes

Advertisement

5-min tech signal

Weekday briefing for engineers who skip the noise.

No spam · Unsubscribe anytime

Advertisement

✈️ CareerPilot

Your AI job-search copilot

Match your resume against live Ashby, Greenhouse & Lever openings — fit scores, job-specific resume optimization and email alerts.

Find matching jobs →

Free Tools

Browse all tools →