Reliable AI agents—even on small,local models.
An open-source agentic harness that prioritizes small local models for reliable, verifiable task completion.
A 90-second film about Agent8088, an open-source agentic harness for small local models, with narration and on-screen text but no captions. It opens on the ways small models fail on long tasks: looping, overflowing their context, or saying done without checking. A research scene notes that at 95% per step a 20-step task finishes 36% of the time, and that in one study results varied 7.8 times more with the harness than with the model. It then shows the Agent8088 command-line start screen with 59 tools and 39 skills, context budgeting and compaction at 75%, recovery from cut-off replies and repeated tool calls, tool selection, permission prompts and sandboxing, and checks that label work as tested, exercised, inspected, unchecked or failed. It ends on the install command and links.
hover for soundAgent8088
The small-local-model-first open-source agentic harness for reliable and verified task completion.
Resolves real pain points through research-backed approaches.
Transparent, sandboxed, safe, in your control. Execution grounded. Robust and context aware.
The installer provisions an isolated Python environment and installs the global agent8088 command. No administrator access required for the base install.
Why Agent8088finishes the task.
Resolves real pain points through research-backed approaches. Transparent, sandboxed, safe, in your control. Execution grounded. Robust and context aware.
Small Models First
- Small-model-first design
- Lean prompts and minimal tool schemas
- Local-model support and model-aware routing
Context That Holds Up
- Robust context management
- Token budget checked before every call, with overflow recovery
- Compaction that keeps your instructions word for word
Smart Tool Use
- Tool intelligence
- Right tools chosen for each request
- Forgiving parsing and checked arguments
Grounded in Real Results
- Grounded execution and verification-gated completion
- Every step judged by real files and test results
- Checks before any “done”, with honest verification labels
Always Moving Forward
- Reliable control loop
- Repeat and stall detection
- Escalates to a stronger model only when needed
Recoverable by Design
- Failure recovery
- Rollback of failed file changes
- Resumable tasks
Sandboxed and Safe
- Transparent, safe and controlled
- Layered permissions and an OS-level sandbox
- Usage monitoring and transparency
Open and Extensible
- Low-friction and extensible
- Terminal, web UI and chat apps
- MCP, skills and sub-agents
Agentic Harness for Making Small Local Models Reliable Agents
TLDR: Small models mostly fail in specific ways. Their context overflows without warning, their tool calls are malformed or wrongly chosen, they say they are done when they are not, and long runs end up looping or losing progress.
Read the post