Capture what went wrong
Execution traces give the teacher context about failed steps and recurring patterns.
An open-source coding-agent harness
Give local agents
routines they
can rely on.
Manifest turns lessons from failed coding-agent runs into reusable, tested routines for small local models.
Explore the prototype ↗From failure to routine
01 Local agent execution
↓ execution traces
02 Frontier teacher analysis
↓ proposed routine
03 Validation & regression gates
↓ accepted change
04 Sandboxed local execution
A research workflow, implemented in the public repository.
The approach
Learn the process.
Keep execution local.
Small models can repeat the same process mistakes across coding tasks. Manifest explores whether a stronger teacher can turn those failures into explicit control-flow routines that a local agent can reuse.
Check before admitting changes
Proposed routines pass schema and code validation, repair, and regression checks before joining the harness.
Constrain what routines can do
The Warden uses a permission manifest and a kernel sandbox to constrain generated routines. Accepted routines can run with a local model through Ollama.
In development
A working prototype.
A clear next experiment.
The public project includes a CLI, a GUI, and an evaluation pipeline. The current teacher uses DeepSeek. We plan to integrate Claude for trace analysis and routine generation, then compare task success, regressions, inference cost, and sandbox compliance using fixed task splits and budgets.
Manifest is an early-stage research prototype. Claude integration is planned; production readiness and performance gains have not been established.
About Manifest
Built by Arshdeep Singh.
Manifest is a bootstrapped venture based in India, founded in October 2026. We are exploring developer tooling that makes small local coding agents more dependable through tested, reusable routines.