A2a Tau Runbook

a2a with tau — Runbook & Caveats

Overview

tau is a non-interactive Zig AI CLI designed for agents. When using tau with a2a for multi-agent coordination, there are specific considerations, caveats, and best practices to ensure successful agent-to-agent communication.

Key Differences from Other CLIs

Critical Configuration for a2a

Required a2a-spawn Flags

# Minimum required flags for tau with a2a:
--no-stream            # MUST disable streaming for tool-calling
--tools bash           # Enable bash tool for a2a coordination (override with TAU_TOOLS)
--max-iterations 40    # MUST increase from default 10 for multi-exchange debates
--timeout-ms 300000    # 5-minute HTTP timeout (slow models need it)
--append-system-prompt $KIT  # Pass kit prompt as system prompt
-p "Begin."            # Single-shot mode

# Optional debugging flags (set via env vars):
TAU_THINKING=1         # Enable --thinking (show reasoning chunks in logs)
TAU_DEBUG=1            # Enable --debug (perf stats + tool I/O to stderr)

Why These Flags Matter

--no-stream: Streaming mode (default) outputs token-by-token, which breaks tool-calling. Tool-calling requires the full response to parse tool_calls before executing them.

--tools bash: tau has 7 built-in tools (bash, read, write, edit, ls, grep, find). For a2a coordination, only bash is needed. Limiting to bash reduces token usage and prevents the model from calling file operations that could interfere with a2a coordination. Override with TAU_TOOLS env var (e.g., TAU_TOOLS="bash,read" to let the agent read files).

--thinking (via TAU_THINKING=1) : Enables reasoning/thinking chunk visibility in logs. Useful for debugging agent behavior. Off by default to save tokens.

--debug (via TAU_DEBUG=1) : Logs performance stats and tool I/O to stderr (visible in agent log). Invaluable for troubleshooting stalled or misbehaving tau agents.

--max-iterations 40: The default agentic loop limit is 10 iterations. A multi-exchange a2a debate requires ~2 tool calls per round (send + recv). For a 3-round debate, you need at least 6 iterations. 40 provides headroom for longer debates and error recovery.

Kit Prompt Best Practices

1. Plain ASCII Messages

Problem: Quotes, apostrophes, and special characters in a2a messages break shell quoting when sent via bash "a2a send ...".

Solution: Restrict messages to plain ASCII without quotes or apostrophes.

Bad kit prompt:

Send your theories to agent-skeptic. Each theory should be detailed and include evidence.

Good kit prompt:

Send your theories to agent-skeptic. Keep each theory to one plain ASCII sentence without quotes or apostrophes.

2. Explicit Tool Instructions

tau's model (xiaomi mimo-v2.5 by default) needs explicit instructions about when and how to use tools.

Required in every tau kit:

IMPORTANT: You have access to the bash tool. Use it to run a2a commands:
- Run: bash "A2A_PROJECT=mars-debate /home/jarancibia/ai/a2a-skill/a2a send agent-skeptic 'your message' --from agent-optimist"
- Run: bash "A2A_PROJECT=mars-debate /home/jarancibia/ai/a2a-skill/a2a recv --as agent-optimist --wait 30"
- Run: bash "A2A_PROJECT=mars-debate /home/jarancibia/ai/a2a-skill/a2a status done --as agent-optimist"

Do NOT call any other tools. Only use bash to run a2a commands.

3. Fixed Round Limits

Problem: Open-ended debates can run indefinitely, burning tokens and budget.

Solution: Specify a fixed number of rounds with a hard cap.

Example:

Debate protocol:
1. Send your first theory
2. Receive counter-argument
3. Send your second theory
4. Receive counter-argument
5. Send your third theory
6. Receive counter-argument
7. Mark yourself as done

Hard cap: 3 rounds max. After 3 empty receives, mark yourself done.

4. Empty Recv Bail-Out

Problem: If one agent finishes early, the other can hang waiting for messages that never come.

Solution: After N empty receives, mark the agent as done.

Example:

Hard cap: 3 rounds max. After 3 empty receives, mark yourself done.

Common Pitfalls

Pitfall 1: Forgetting --no-stream

Symptom: Agent hangs, no tool calls executed, log shows streaming output.

Fix: Always include --no-stream in a2a-spawn for tau.

Pitfall 2: Default --max-iterations Too Low

Symptom: Agent stops mid-debate after 2-3 rounds, doesn't complete protocol.

Fix: Use --max-iterations 40 or higher for multi-exchange coordination.

Pitfall 3: Shell Quoting Failures

Symptom: a2a send fails with syntax error, message never reaches bus.

Fix: Restrict kit prompt messages to plain ASCII without quotes/apostrophes.

Pitfall 4: Model Doesn't Recognize Tool Instructions

Symptom: Agent responds to kit prompt but doesn't execute bash commands.

Fix: Make tool instructions explicit and prominent in kit prompt. Use "IMPORTANT:" prefix.

Pitfall 5: Agent Never Marks Done

Symptom: Agent stays status=active indefinitely, process doesn't exit.

Fix: Include explicit "Mark yourself as done" instruction in protocol. Consider manual intervention if needed.

Provider-Specific Notes

xiaomi (mimo-v2.5) - Default

• Pros: Fast, cost-effective, good tool-calling
• Cons: May need more explicit instructions than other models
• Notes: Works well with a2a when kit prompts are clear

openai (gpt-4o-mini)

• Pros: Excellent instruction following, robust tool-calling
• Cons: More expensive, slower
• Notes: May require different prompt structure

deepseek (deepseek-chat)

• Pros: Good value, capable tool-calling
• Cons: May have different output format expectations
• Notes: Test thoroughly before production use

Debugging Tips

1. Check Agent Logs

tail -f /tmp/a2a-{agent}.log

Look for:

• Streaming output (indicates --no-stream missing)
• Tool call parsing errors
• Bash command execution failures

2. Monitor the Bus

a2a peek --limit 10 --project {project}

Check if messages are landing on the bus. If not, the bash tool isn't working.

3. Check Agent Status

a2a list --json --project {project}

Verify:

• Both agents are registered
• PIDs are correct
• Status transitions (active → done)

4. Test Tool-Calling in Isolation

./zig-out/bin/tau --tools bash --no-stream -p "Run: bash 'echo test'"

If this fails, the issue is with tau's tool-calling, not a2a.

Example: Working Smoke Test

See the tau repository smoke.md for a complete working example:

• Project: mars-debate
• Agents: agent-optimist, agent-skeptic
• Protocol: 3-round debate
• Kit prompts: Plain ASCII, explicit tool instructions

Integration Checklist

Before deploying tau with a2a:

• a2a-spawn updated with --no-stream --tools bash --max-iterations 40 --timeout-ms 300000
• TAU_TOOLS, TAU_THINKING, TAU_DEBUG env vars configured if needed
• Kit prompts use plain ASCII (no quotes/apostrophes)
• Kit prompts have explicit tool instructions
• Kit prompts specify fixed round limits
• Kit prompts include empty recv bail-out
• tau binary built and accessible (./zig-out/bin/tau or in PATH)
• a2a project initialized
• Agents registered before spawning
• PIDs updated after spawn with --upsert
• Monitoring plan in place (peek, list, logs)

Performance Considerations

Token Usage

• tau uses JSON output by default (more tokens than text)
• Tool-calling adds overhead (schemas, tool_calls, results)
• Consider --mode text for non-a2a use cases

Process Management

• tau processes are backgrounded via nohup in a2a-spawn
• Processes exit when agent marks done or max iterations reached
• Manual kill may be needed if agent hangs

Resource Limits

• tau is a compiled Zig binary (~250KB), very lightweight
• No Node.js runtime required
• Minimal memory footprint

Future Enhancements

Potential improvements for tau + a2a:

1. Native a2a client: Direct SQLite access instead of bash wrapper
2. Config file support: a2a settings in ~/.config/tau/config.json
3. Provider-specific defaults: Different max-iterations per provider
4. Automatic done detection: Heuristics to detect completion
5. Better error messages: Clearer tool-calling failure diagnostics

References

• tau README: https://github.com/javimosch/tau
• a2a-skill repo: https://github.com/javimosch/a2a-skill
• a2a AGENTS.md: Full repo guide and protocols
• a2a SMOKE_TEST.md: General smoke testing patterns