A2A Agent Orchestration
Multi-agent workflows — chain focused LLM calls instead of one mega-prompt.
A single prompt that "does everything" is brittle. We split the work into focused agents (sourcing → screening → coordinator), each with its own job, prompt, and output schema. Every step is audited, replayable, and inspectable — including model, latency, and tokens.
What you get
Built to solve the actual problems hiring teams hit every day.
Focused agents, not one mega-prompt
Each step has a single job. Sourcing strategist defines the ideal candidate; screening agent scores the pool; coordinator synthesizes a recommendation.
Full audit trail
hr_agent_runs persists every step — agent, role, input, output, latency, model, tokens. Replay or inspect any orchestration.
Two workflows shipped
"Source → Screen → Match" runs the full hiring chain. "Requirements → JD" turns a goal into a polished job description.
Cost & latency visible
Total tokens and milliseconds are tracked per run and per step — no more guessing how much an AI workflow cost.
Extensible
Add a new workflow by appending one case to the orchestrator switch — the audit, rate limit, and UI all come free.
How it works
Four steps from setup to value.
Pick a workflow
On /ai-agents click "Launch workflow", choose source_screen_match or reqs_to_jd.
Provide context
Source-screen-match needs a position; reqs-to-jd needs a goal. Both accept an optional goal text.
Agents run sequentially
Each agent's output becomes the next agent's input. The chain typically takes 10–30 seconds.
Inspect every step
Expand any run to see each agent's prompt-input, raw output, latency, and token count. The final result is rendered separately.
FAQ
Quick answers to common questions.
Why split into multiple agents instead of one big prompt?
Focused prompts are more reliable, easier to debug, and let each step use a different model if needed. The trade-off is more API calls — we surface latency and tokens so you can see the cost.
Are the workflows configurable?
The two shipped workflows are server-defined for safety (no arbitrary prompt construction from the client). Adding new ones is a one-case change in the orchestrator.
Does this need GovCon?
No — orchestration is on every Pro and GovCon plan. The "ai" entitlement gates it.