July 2, 2026
Conductor, higher free-tier token limits, and more
Conductor is live, and it's a big one. Here's what else shipped this week.
Retell's copilot for building and maintaining voice agents. Three parts of it worth knowing about:
Ask Conductor to run a scenario, like "what happens if the caller wants to reschedule," and it plays that call against your agent without dialing a real number, then shows you exactly how it responded. Every proposed change is validated against the platform first, so nothing reaches your agent unless it actually holds up.
→ Simulate a specific scenario on demand, no real call required
→ Every change validated against the platform before you ever see it
→ Catches edge cases most teams only find after a call goes wrong
→ Nothing ships until the graph, data connections, and logic all check out
Tested before it's trusted.
Most bad calls just sit in a call log until someone happens to notice a pattern. Conductor reads failed calls, writes a test that reproduces the problem, runs it as a simulation without dialing a real number, and proposes the fix.
→ Failed calls turned into test cases automatically
→ Simulated before anything touches production
→ Fixes proposed as a change you review, not applied silently
→ Already handles 70% of Retell's own simulation testing internally
Every bad call becomes the reason the next one goes better.
Conductor never edits a live agent on its own. Every change it proposes shows up inside the actual agent flow, on the exact setting it affects, so you can see precisely what would change before it does.
→ Changes appear in place, not as a JSON diff
→ Side-by-side view of the old and new value
→ Accept or reject one change at a time, with undo
→ Nothing goes live until you approve it
The AI suggests the change. You still decide.
→ Free LLM token allowance raised to 4,000: every agent now gets 4,000 free LLM tokens before any surcharge applies, and chat agents no longer show a surcharge section at all since it doesn't apply to them.
→ Better KB parsing and chunking: documents parse more accurately now, especially tables, headers, and unusual layouts, and each chunk keeps its parent header info and splits at sentence boundaries instead of mid-word, so your agent pulls cleaner context.
