Darkmind / Issue 038

Draw the Dependency Line Before You Parallelize Agents

Split only the work that can finish from the same stable input, then give each parallel agent a bounded deliverable and a single synthesis rule.

Nov 14, 20264 min readAgent design

Parallel agents can make a workflow look faster while making it less reliable. A coordinator sees a request such as “compare these vendors and recommend one,” launches four researchers, then waits. But if one researcher needs the compliance constraints found by another, the team has not parallelized independent work. It has started a race between incomplete interpretations. The eventual synthesis may look comprehensive while quietly comparing different requirements, versions, or scopes.

The practical move is to draw the dependency line before you fan out. Put everything that must be shared, chosen, or validated first into a short common brief. Then launch parallel workers only for subproblems that can finish from that same frozen brief without needing another worker’s output. Give each worker a bounded artifact—not “research vendor X”—and give the final synthesizer a rule for handling disagreement and missing evidence.

OpenAI recommends multi-agent work when research, analysis, or implementation can be split into concrete, bounded, independent workstreams; it cautions that subagents increase token use and that ordered chains are better kept with one agent.[1] Google makes the same distinction in architectural terms: its parallel pattern is for independent concurrent subtasks whose outputs are synthesized, while its sequential pattern fits structured work where one output directly feeds the next.[2] The key question is not “can several agents work at once?” It is “would a worker’s correct answer change if it saw another worker’s result?” If yes, put that dependency before or after the parallel stage.

Find the shared brief first

Start with the facts every worker must interpret the same way: the decision to support, scope, date or version boundary, evaluation criteria, allowed sources or tools, and a definition of done. This common brief is not a long prompt. It is a small contract that prevents each worker from inventing its own assignment.

For a vendor comparison, the brief might pin the target use case, geography, data classification, budget range, and evidence cutoff date. A security reviewer, cost analyst, and integration analyst can then work independently because they share the criteria. But do not launch a recommendation writer in parallel with those reviewers. Its job depends on their findings. Likewise, if an agent needs to choose which repository, customer, or policy applies, resolve that target before parallel work begins.

A useful test is the late-arrival test: imagine one worker finishes last. Would its evidence require any other worker to redo a material conclusion? If the answer is yes, the work is coupled. Either move the prerequisite into the shared brief or make the dependent step sequential.

Copy this parallel-work card

Use this before allowing a coordinator to launch subagents or parallel tool branches.

COMMON BRIEF (frozen before fan-out)
- decision / question: [one observable outcome]
- scope: [tenant, project, corpus, date/version boundary]
- success criteria: [how the final result will be judged]
- shared evidence rules: [eligible sources, required citations, exclusions]
- budget: [max workers, time, tokens/cost]

DEPENDENCY CHECK
For each proposed worker, answer:
- Can it complete using only COMMON BRIEF + its assigned inputs? yes | no
- Does it need another worker's finding to choose its method, scope, or claim? yes | no
- Could its output change another worker's factual conclusion? yes | no
If any answer requires another worker's output, make that edge sequential.

PARALLEL WORKER CARD
- worker_id: [stable name]
- owned question: [one non-overlapping subproblem]
- inputs: [exact files, URLs, IDs, or bounded query]
- return schema: {claim, evidence_refs, caveats, status: complete|partial|blocked}
- must not: [choose final answer, broaden scope, call writes]

SYNTHESIS RULE
- accept only claims with required evidence
- label partial or conflicting findings; do not average them into certainty
- resolve conflicts by [named criterion or a fresh targeted check]
- final outcome: recommend | compare | insufficient_evidence

The owned question line is where duplication becomes visible. “Find useful facts about Vendor A” and “research Vendor A’s risks” will produce overlapping searches and competing summaries. Better assignments are “extract published data-residency commitments from the named documents” and “check the integration requirements against the stated stack.” Each output has a different evidence shape and can be checked without reading a second worker’s scratchpad.

Keep synthesis out of the race

Parallel work needs a gather step with its own discipline. Collect results in a common schema, preserve worker IDs and source references, and distinguish complete from partial or blocked. Do not tell a synthesizer that all agents “researched the topic”; tell it precisely which question each answered and which one did not.

Then make synthesis a sequential decision. It may compare evidence, identify a gap, or send one targeted follow-up. It should not silently fill a missing security finding with a cost worker’s optimistic wording. Google notes that parallel systems add cost and complexity, including the need to synthesize potentially conflicting outputs.[2] In practice, the synthesis rule is the difference between useful parallelism and several independent drafts pasted together.

Failure mode: parallelizing the uncertainty

More workers do not turn an unresolved premise into a fact. Launching five agents before you know which policy version applies simply produces five answers built on potentially different assumptions. Parallelism can also multiply access, cost, and noisy duplicate retrieval.

Keep fan-out narrow and evidence-bounded. Resolve targets, permissions, and shared criteria first; run dependent reasoning in order; and stop when a worker reports a missing prerequisite rather than asking it to guess. For short tasks or one ordered chain, a single agent is often the cheaper and more coherent design.[1]

The habit to keep

parallelize independent evidence work, not decisions that depend on one another. Freeze the shared brief, draw the dependency line, and make one final stage responsible for reconciling what the workers actually proved.

Sources

  1. OpenAI, “API deployment checklist”
  2. Google Cloud, “Choose a design pattern for your agentic AI system”