Caveman

Built for agents Caveman Cloud

Your agent.
Its own
feedback loop.

Let your coding agent learn from its work. Read traces, find the friction, and test better instructions. All from the terminal it already calls home.

Look inside

Your coding agent. Your repository. Better next runs.

Watch an agent improve its own workflowInteractive example
coding-agent~/my-project
Caveman Cloud Project connected

Find repeated work in my last coding session.

I’ll look for repeated tool calls in Caveman Cloud.

caveman_readtraces.search
↳ 24 traces found session_014
TraceActivityCalls
trace_080Explore repository4
trace_082Read instructions ×38
trace_085Run test suite2
One pattern worth a closer look.
improve/instructionsReading session history

Illustrative session and results. Real Cloud operation names. Changes stay reviewable.

A shared language for your agent and your platform. CLI MCP Structured data

Context it can work with

Your platform.
In its language.

Traces, datasets, evaluations, and scenarios are objects your agent can read and act on. Every investigation starts with the evidence behind a run.

Discover available operations Inspect exact input schemas Work within project permissions
Explore the platform
Caveman CloudStructured. Queryable.
Select a resource
to look inside.
trace.jsonResource excerpt
{  "trace_id": "trace_082",  "session_id": "session_014",  "agent_slug": "coding-agent",  "request_count": 8,  "error_count": 0,  "input_tokens": 18420,  "cached_input_tokens": 6120}
caveman_read → traces.getExample data
The steps behind the outcome.

Follow a session through its requests, tool calls, errors, and token usage. Give the agent evidence it can investigate.

Improvement that stays with you

A better run starts
with a change you can read.

Let the agent update its instructions and test the difference. Keep the diff, the evidence, and the decision in your workflow.

01 / Find the pattern
read_file AGENTS.mdread_file AGENTS.mdread_file AGENTS.md

Remember what went wrong.

Find repeated work across the session, then trace it back to the instruction that caused it.

02 / Change the instructions
AGENTS.md− Re-read before every tool call.+ Read once. Refresh on change.

Make a focused edit.

Give the next run a better starting point through a versioned change in your own repository.

03 / Keep the evidence
Same cases. Quality holds.Diff + evaluation results

Earn the next version.

Compare the candidate against the baseline. Review the result before adopting the change.

From one run to the next

Give your agent
something to learn from.

Bring your coding agent and a workload worth improving. We’ll help you connect the loop.

Private preview · Project-scoped access