Improve Agents

Build, improve, and iterate on agents using Claude Code and evals.

Your platform is running. Now make it yours.

The templates are designed for coding agents. Claude Code can read traces, run evals, and edit agent code in a tight feedback loop. Three of the five workflows run autonomously with no input from you.

Claude Code Workflows

The docs/ directory includes five prompts that cover the agent development lifecycle:

Prompt	What it does	Autonomous
create-new-agent.md	Scaffolds an agent, registers it, smoke-tests via cURL	No
improve-agent.md	Derives probes from instructions, runs them, fixes failures	Yes
extend-agent.md	Add tools or features. You direct, Claude executes	No
hill-climb.md	Runs evals, diagnoses failures, fixes until all pass	Yes
review-codebase.md	Finds drift between docs, code, and config	Yes

Run any prompt by pasting it into a Claude Code session:

Run docs/create-new-agent.md

Evals

Lock in agent behavior with the eval suite:

python -m evals                # run the suite
python -m evals -v             # verbose with rich panels
python -m evals --case <name>  # run one case

Evals use AgentAsJudgeEval (LLM judge, binary pass/fail) and ReliabilityEval (tool-call assertions). Results log to Postgres so you can track behavior over time.

The hill-climb.md prompt runs evals, diagnoses failures, and fixes what’s in scope. Stops when all cases pass.

Agent Patterns

The templates ship with two agents that demonstrate different patterns:

Agent	Pattern	Description
WebSearch	Direct tools	Agent sees each tool individually
CodeSearch	Context provider	Agent sees one query_<thing> tool that hands off to a sub-agent

Direct tools work when the agent needs fine-grained control over each tool call. Context providers work when you want to encapsulate a capability (like searching a codebase) behind a single interface.

Copy either pattern when building your own agents.

Teams and Workflows

For most things, one agent is enough. When it isn’t:

Pattern	When to use
Multi-agent teams	Route to specialists, coordinate parallel work, synthesize results
Agentic workflows	Deterministic pipelines that run the same way every time

Rule of thumb: agents for open questions, teams for routing, workflows for processes.

Scheduled Tasks

scheduler=True is on by default. Schedule any agent or workflow on a cron:

• Maintenance. Purge sessions older than 90 days. Vacuum tables.
• Proactive runs. Every weekday morning, summarize overnight news and send to Slack.
• Periodic re-evaluation. Run the eval suite on cron to catch behavior drift before users do.

See Scheduler docs for the cron API.

Interfaces

Agents should live where your users are.

Slack is pre-wired in the templates. Set SLACK_BOT_TOKEN and SLACK_SIGNING_SECRET in your env and the interface activates automatically. See Slack setup.

For Discord, Telegram, WhatsApp, or custom UIs, see the Interfaces guide.

Next Steps

Tools Catalog

100+ integrations ready to add to your agents.

AgentOS Features

Sessions, memory, tracing, and more.