Safa Global · Claude Code
4.8 + ultracode + Workflows. The release that turns Claude Code from a turn-by-turn pair-programmer into a background orchestrator of hundreds of agents, and what it means for how Safa ships work.
Executive audio briefing. Two hosts, about 22 minutes.
The takeaway
Opus 4.8 is not mainly a "smarter model" story. The real shift is dynamic workflows: Claude now writes a script that orchestrates dozens to hundreds of subagents in the background, with agents adversarially checking each other, while your session stays free. Work scoped in quarters can finish in days.
ultracode (/effort ultracode) is the easy button: maximum reasoning effort plus automatic workflow orchestration, so Claude decides when a task deserves an agent swarm. Use it for codebase-wide audits and large migrations; keep routine work on /effort high - workflows cost meaningfully more tokens.
Anthropic's most capable generally-available model, released 28 May 2026. Built for professional software engineering, complex agentic work, and high-stakes enterprise tasks - at the same token price as 4.7.
high (the 4.8 default) → xhigh ("extra") → max. Higher effort means Claude thinks more often and more deeply, trading tokens for quality; rate limits are raised to compensate.| Lever | Detail |
|---|---|
| Standard price | $5 / M input · $25 / M output (unchanged from 4.7) |
| Fast mode | $10 / M input · $50 / M output - 3× cheaper than on prior models |
| Caching / batch | Up to 90% off with prompt caching · 50% off with batch |
| Effort levels | high (default) · xhigh · max |
A dynamic workflow is a JavaScript script Claude writes that orchestrates subagents at scale. A runtime executes it in the background while your chat session stays responsive, and only the final answer lands back in Claude's context.
The key mental shift is who holds the plan. With subagents and skills, Claude is the orchestrator and every intermediate result fills its context. A workflow moves the loop, the branching, and the intermediate results into code - so it can apply a repeatable quality pattern: independent agents draft a plan from several angles, or adversarially review each other's findings, before anything is reported.
| Subagents | Skills | Workflows | |
|---|---|---|---|
| What it is | A worker Claude spawns | Instructions Claude follows | A script the runtime runs |
| Who plans | Claude, turn by turn | Claude, per prompt | The script |
| Results live in | Claude's context | Claude's context | Script variables |
| Scale | A few per turn | Same | Dozens to hundreds |
| If interrupted | Restarts the turn | Restarts the turn | Resumable |
Up to 16 concurrent agents (fewer on low-core machines), a hard cap of 1,000 agents per run, no mid-run user input (run each sign-off stage as its own workflow), and resume works only within the same Claude Code session.
Proof point: Bun was ported from Zig to Rust - ~750,000 lines, 99.8% of the existing test suite passing - in eleven days using dynamic workflows, with two reviewer agents per file.
ultracode = xhigh reasoning effort + automatic workflow orchestration. Turn it on and Claude plans a workflow for every substantive task itself, instead of waiting for you to ask.
| Enable | /effort ultracode - lasts the current session, resets on a new one |
| Drop back | /effort high when you return to routine work |
| Behaviour | One request can become several workflows in a row: understand the code → make the change → verify it |
| The cost | Every task uses more tokens and takes longer. Only available on models that support xhigh effort |
Think of the three gears: /effort high for everyday work, the workflow keyword for a single big task on demand, and /effort ultracode when you're heads-down on something large and want Claude to escalate on its own.
/deep-research <question> fans out web searches across angles, cross-checks sources, votes on each claim, and returns a cited report with weak claims filtered out. The fastest way to see a workflow run.workflow anywhere in your prompt and Claude writes a script instead of grinding turn-by-turn. Press alt+w to cancel the trigger if you didn't mean it./effort ultracode and let Claude decide per task./workflowsRuns happen in the background; /workflows opens the progress view - phases with agent counts, token totals, and elapsed time.
| Key | Action |
|---|---|
| p | Pause / resume the run |
| x | Stop an agent, or the whole workflow |
| r | Restart a running agent |
| s | Save the run's script as a reusable /command |
Save to .claude/workflows/ (shared with the repo) or ~/.claude/workflows/ (personal, every project). A saved workflow becomes /<name> in future sessions - the orchestration itself becomes repeatable, not just the prompt.
Research preview. Needs Claude Code v2.1.154+. On by default for Max and Team; turn on from the Dynamic workflows row in /config on Pro; Enterprise needs admin activation. Works in CLI, Desktop, IDE extensions, claude -p, and the Agent SDK.
| Reach for a workflow | Stay in normal chat |
|---|---|
| Codebase-wide audits (security, dead code, missing auth) | A single-file edit or quick fix |
| Large migrations: framework swaps, API deprecations, language ports across hundreds of files | A question you can answer in one turn |
| High-stakes work where a wrong answer is costly and you want adversarial verification | Anything one conversation already handles well |
| Research that must cross-check many sources against each other | Exploratory back-and-forth where you steer each step |
A run spawns many agents, so it can use meaningfully more tokens than doing the same task in conversation, and it counts toward your plan's usage and rate limits. Start scoped to learn the usage pattern. Check /model before a large run, and ask Claude to route stages that don't need the strongest model to a smaller one.
Where ultracode and workflows earn their token cost first, given our stack:
~/aquiii-webapp/src for missing auth checks, then verify the pending Shopify token rotation is fully propagated. Independent agents per route, adversarially reviewed.SKILL.md frontmatter and cross-links in one run, instead of editing them one by one./branch-review workflow. Build the review once on a feature branch, press s, and reuse the same multi-agent orchestration on every branch from then on./effort high day-to-day; reserve /effort ultracode for migrations and audits where the extra tokens buy verified correctness.